application of monitoring to dynamic characterization and …745269/fulltext01.pdf · in a steel...

Application of Monitoring to Dynamic Characterization and Damage Detection in Bridges

IGNACIO GONZALEZ

Doctoral Thesis

Stockholm, Sweden 2014

Akademisk avhandling som med tillstånd av Kungliga Tekniska högskola framlägges till offentlig granskning för avläggande av teknologie doktorsexamen i Brobyggnad fredagen den 19 september 2014 kl 10:00 i sal F3, Kungliga Tekniska högskola, Lindstedsvägen 26, Stockholm.

© Ignacio Gonzalez, September 2014

Tryck: Universitetsservice US-AB

TRITA-BKN. Bulletin 126, 2014 ISSN 1103-4270 ISRN KTH/BKN/B-126-SE

i

Preface

This Doctoral thesis was carried out at the Department of Civil and Architectural Engineering, the division of Structural Engineering and Bridges, at KTH Royal Institute of Technology in Stockholm, during 6 years of research with a 80% dedication. My most deep gratitude is expressed to Professor Raid Karoumi, who supervised this thesis and was the real driving force behind it. Thanks to Costin Pacoste for being my co-supervisor and Jean-Marc Battini for reviewing this work.

Thanks also to all my colleagues and friends at the Division of Structural Engineering and Bridges. Especially to laboratory technicians Stefan Trillkott and Claes Kullberg who managed the instrumentations in several of the studies.

Many thanks to Lärkstaden and to all those who have passed through there these years, making Stockholm home. Thanks to my family and friends for their invaluable encouragement.

Stockholm, September 2014

Ignacio González

iii

Abstract

The field of bridge monitoring is one of rapid development. Advances in sensor technologies, in data communication and processing algorithms all affect the possibilities of Structural Monitoring in Bridges. Bridges are a very critical part of a country’s infrastructure, they are expensive to build and maintain, and many uncertainties surround important factors determining their serviceability and deterioration state. As such, bridges are good candidates for monitoring. Monitoring can extend the service life and avoid or postpone replacement, repair or strengthening works. The amount of resources saved, both to the owner and the users, by reducing the amount of non-operational time can easily justify the extra investment in monitoring.

This thesis consists of an extended summary and five appended papers. The thesis presents advances in sensor technology, damage identification algorithms, Bridge Weigh-In-Motion systems, and other techniques used in bridge monitoring. Four case studies are presented. In the first paper, a fully operational Bridge Weigh-In-Motion system is developed and deployed in a steel railway bridge. The gathered data was studied to obtain a characterization of the site specific traffic. In the second paper, the seasonal variability of a ballasted railway bridge is studied and characterized in its natural variability. In the third, the non-linear characteristic of a ballasted railway bridge is studied and described stochastically. In the fourth, a novel damage detection algorithm based in Bridge Weigh-In-Motion data and machine learning algorithms is presented and tested on a numerical experiment. In the fifth, a bridge and traffic monitoring system is implemented in a suspension bridge to study the cause of unexpected wear in the bridge bearings.

Some of the major scientific contributions of this work are: 1) the development of a B-WIM for railway traffic capable of estimating the load on individual axles; 2) the characterization of in-situ measured railway traffic in Stockholm, with axle weights and train configuration; 3) the quantification of a hitherto unreported environmental behaviour in ballasted bridges and possible mechanisms for its explanation (this behaviour was shown to be of great importance for monitoring of bridges located in colder climate) 4) the statistical quantification of the non-linearities of a railway bridge and its yearly variations and 5) the integration of B-WIM data into damage detection techniques.

Keywords: Structural health monitoring, Traffic monitoring, Bridge monitoring, Bridge Weigh-In-Motion, BWIM, Damage detection, Suspension bridge bearings, Axle loads, Dynamics, Temperature effect.

v

Sammanfattning

Broövervakning är ett område under snabb utveckling. Framsteg inom sensorteknik, datakommunikation och algoritmer för databehandling möjliggör tillståndsbedömning, skadeidentifiering, trafikövervakning och andra tillämpningar av övervakningssystem. Broar är en vital del av vår infrastruktur, de utgör stora kostnader avseende såväl byggnation som underhåll, samtidigt som osäkerheterna är stora avseende dess brukstillstånd samt nedbrytningsprocesser. Detta gör broar till lämpliga objekt för övervakning. Genom övervakning kan livslängden förlängas, varvid man kan undvika eller senarelägga ett utbyte eller förstärkningsåtgärder. Många broar utgör även flaskhalsar i transportsystemet med få eller inga alternativa transportvägar. De resurser som kan sparas genom att minska trafikstörningarna, för både anläggningsägare och användare, kan enkelt rättfärdiga investeringskostnaden för övervakningssystemen.

Föreliggande uppsats består av en utökad sammanfattning samt fem bifogade artiklar. I uppsatsen presenteras framsteg inom sensorteknik, algoritmer för skadeidentifiering samt metoder för övervakning av trafiklaster genom mätning på broar, Bridge Weigh-In-Motion. Artiklarna baseras på fyra fallstudier. I den första artikeln har ett fullt fungerande Bridge Weigh-in-Motion-system utvecklats, vilket har tillämpats på en järnvägsbro av stål. Insamlad data har analyserats för att erhålla objektspecifika uppgifter om trafiklaster. I den andra artikeln karakteriseras den årliga variationen i styvheten hos en samverkansbro, med hänsyn till dess naturliga spridning. I den tredje, studeras de icke-linjära egenskaper hos en samverkansbro. De beskrivs stokastiskt, med hänsyn till deras årliga variation. I den fjärde, en ny algoritm för skadedetektering baserad på BWIM-data samt maskininlärningstekniker presenteras och testas i ett numeriskt experiment. I den femte, har ett bro- och trafikövervakningssystem implementerats på en hängbro, i syfte att undersöka orsaken till oväntad nedbrytning av brolagren.

Några av arbetets viktigaste vetenskapliga bidrag är: 1) utvekling av ett B-WIM system för tågtrafik som kan uppskatta laster på enstaka axlar, 2) in-situ beskrivning av trafik på en järnvägsbro med, bland annat, axellaster och tågkonfigurationer 3) upptäkten och kvantifiering av hittills opublicerade temperaturrelaterade förändringar i dynamisk beteende hos ballasterade broar, av stor vikt vid övervakning av broar i kalla regioner, 4) statistisk beskrivning av olinjäriteter på en järnvägsbro, och dess årliga variationer med temperatur samt 5) integrering av B-WIM data med skadedetekteringtekniker.

Nyckelord: Tillståndsbedömning genom övervakning, Trafikövervakning, Bridge Weigh-In-Motion, BWIM, Broövervakning, Skadeidentifiering, Hängbrolager, Axellaster, Dynamik, Temperatureffekter.

vii

List of Publications

Four journal papers and one conference paper form the basis of this Thesis.

Paper I: I. Gonzalez, and R. Karoumi. "Traffic monitoring using a structural health monitoring system." (2013), ICE Bridge Engineering (accepted for publication). http://dx.doi.org/10.1680/bren.11.00046

Paper II: I. Gonzales, M. Ülker-Kaustell, and R. Karoumi. "Seasonal effects on the stiffness properties of a ballasted railway bridge." Engineering Structures 57 (2013): 63-72.

Paper III: I. Gonzalez, and R. Karoumi. "Analysis of the annual variations in the dynamic behavior of a ballasted railway bridge using Hilbert transform." Engineering Structures 60 (2014): 126-132.

Paper IV: I. Gonzalez, and R. Karoumi. "BWIM Aided Damage Detection in Bridges Using Machine Learning. " Submitted to Computers and Structures (August 2014).

Paper V: I. Gonzalez, and R. Karoumi. "Continous monitoring of bearing forces and displacements in the High Coast Bridge." The 7th International Conference on Bridge Maintenance, Safety and Management (IABMAS 2014), Shanghai. Paper 0130P.

Paper I, III, IV and V were planned, implemented and written by the author. In paper II the contribution of the author was the stochastic model updating and data analysis, while the second author performed the finite element modelling and the editing of the paper. The planning of this work was performed in close collaboration by the first and second authors.

Other Publications by the author:

JOURNALS

J. Shu, Z. Zhang, I. Gonzalez, R. Karoumi. 2012. "The application of a damage detection method using Artificial Neural Network and train-induced vibrations on a simplified railway bridge model. " Engineering Structures 52 (2013): 408-421.

CONFERENCES

I. González. 2010. "The Validity of Simplified Dynamic Analysis of the New Årsta Bridge’s Response to Moving Trains." Tenth International Conference on Computational Structures Technologies, Valencia. Paper 25.

L. Mottola, T. Voigt, I. Gonzalez Silva, R Karoumi. "From the desk to the field: Recent trends in deploying Wireless Sensor Networks for monitoring civil structures." Sensors, 2011 IEEE, Limerick, 62-65.

I. Gonzalez, R. Karoumi, A. Llorens. 2012. "Improved bridge Respons evaluation based on dynamic testing." The 6th International Conference on Bridge Maintenance, Safety and Management (IABMAS 2012), Stressa. Paper 1812.

I. Gonzalez. 2013. "Temperature dependance of ballast stiffness on a railway bridge." International IABSE conference, Rotterdam. Paper 262.

I. Gonzalez. 2013. "Stochastic Model Updating of Ballast Stiffness in Cold Conditions." The 6th International Conference on Structural Health Monitoring of Intelligent Infrastructure (SHMII 2013). Hong Kong. Paper 1062.

REPORTS

I. González, R. Karoumi. 2010. "Continuous Monitoring of the High Coast Suspension Bridge. Measurement Period February to December 2010." Technical Report 2011:03, Structural Design and Bridges 2011, ISSN 1404-8450.

ix

Contents Preface ........................................................................................................................................ i

Abstract .................................................................................................................................... iii

Sammanfattning ....................................................................................................................... v

List of Publications ................................................................................................................. vii

1 Introduction ...................................................................................................................... 1

1.1 Background ............................................................................................................. 1

1.2 Aims and Scope ....................................................................................................... 1

1.3 Outline of the Thesis ............................................................................................... 3

2 Structural Health Monitoring......................................................................................... 5

2.1 History of Bridge SHM (applications) .................................................................... 6

2.2 Sensors used in SHM of bridges ............................................................................. 8

2.2.1 Cameras ...................................................................................................... 8

2.2.2 Fibre Optic ................................................................................................ 10

2.2.3 Electrochemical (Corrosion) .................................................................... 13

2.2.4 Laser Doppler Vibrometer ....................................................................... 14

2.2.5 Accelerometers ......................................................................................... 14

2.2.6 Strain & Relative Displacement Sensors ................................................. 14

2.2.7 Temperature sensors ................................................................................. 15

2.2.8 Acoustic emissions ................................................................................... 15

2.3 Algorithms used in SHM of bridges ...................................................................... 16

2.3.1 Data analysis & evaluation ....................................................................... 16

2.4 Other aspects ......................................................................................................... 25

2.4.1 Data communication ................................................................................ 25

2.4.2 Sensor placement ...................................................................................... 27

2.4.3 Sensor failure detection ............................................................................ 28

2.5 Concluding Remarks ............................................................................................. 28

3 Bridge Weigh-In-Motion ............................................................................................... 31

3.1 History of BWIM .................................................................................................. 32

3.2 Recent Algorithms and Applications .................................................................... 33

4 Advanced techniques for system identification and damage detection .................... 37

4.1 Artificial Neural Networks .................................................................................... 37

x

4.2 Gaussian Processes ................................................................................................ 38

4.3 Hilbert Transform .................................................................................................. 41

5 Case Studies .................................................................................................................... 45

5.1 The Söderström Bridge ......................................................................................... 45

5.2 The Skidträsk Bridge ............................................................................................. 47

5.3 The High Coast Bridge .......................................................................................... 48

5.4 Damage detection .................................................................................................. 50

6 Conclusions .................................................................................................................... 53

6.1 General Conclusions .............................................................................................. 53

6.2 Further Research .................................................................................................... 54

Bibliography ........................................................................................................................... 57

A Appendix A – Paper I - ............................................................................................. 67

B Appendix B – Paper II - ............................................................................................ 67

C Appendix C – Paper III - ............................................................................................. 67

D Appendix D – Paper IV - ............................................................................................. 67

E Appendix E – Paper V - ............................................................................................... 67

81 97 103 115

1.1. BACKGROUND

1

1 Introduction

1.1 Background

Railway and highway bridges are an important part of the transport infrastructure. These bridges represent a major investment by society, and a major portion of annual infrastructure management costs go to their inspection and maintenance. Bridges often constitute bottlenecks in the transport system, with few practical alternative routes. As such, closing them for repair, inspection or replacement places high costs on the users. Furthermore, the safety levels in bridges are expected to be higher than in other parts of the transport system because the failure of a bridge can have severe consequences in terms of material damage and human lives. Introducing various monitoring techniques (damage detection, traffic monitoring, reliability assessment, etc.) can save costs by improving understanding of the structure, thus reducing the need for overly safe assumptions and allowing for the possibility of early warnings on problems.

As with all infrastructure, bridges age, and their performance worsens. At the same time, the demands imposed on bridges generally increase with time in the form of faster and heavier traffic. The cost of strengthening a bridge, to the bridge owner, the users and society at large, does not decrease over time. Rather, the cost of labour, materials and traffic interruptions can arguably be said to increase as time passes. On the other hand, the cost of structural health monitoring components, sensors, general computers and networks decreases each year, while the capabilities of structural health monitoring systems are constantly improving with the help of new algorithms and sensors. The economic value of structural health monitoring is a separate issue that has been the subject of intensive studies [1]. In light of these facts, structural health monitoring is likely to become more common and advantageous to bridge owners, reducing the ever-increasing costs of inspection, repair and replacement while reducing hardware and software costs and increasing the structure’s reliability using real time information on the monitored bridge. Thus, the study and development of structural health monitoring techniques will gain relevance and value in the near future.

1.2 Goals and Scope

The goal of this study was to provide practical tools for improving bridge operation and maintenance routines through monitoring. To that end, a more specific goal was to survey the latest developments in structural health monitoring. The survey was limited to recent advances in sensor technology, data processing algorithms and a number of other aspects related to structural health monitoring, with a focus on damage detection.

Chapter 1. Introduction

2

The behaviour of a bridge at any given point in time is determined by three main effects (see figure 1):

1) the traffic, 2) the environmental conditions, and 3) the structural characteristics of the bridge.

The purpose of damage detection is to identify and infer changes in the structural characteristics of a bridge by monitoring its behaviour. To realise a damage detecting algorithm in a real bridge, it becomes necessary to understand the effects that traffic and environmental conditions have on it so that observed changes in the bridge’s behaviour due to these effects are not wrongly attributed to damage.

Traffic: Paper I and, to a lesser extent, paper V address the study and identification of traffic loads and load effects based on data gathered by a bridge monitoring system.

Environmental conditions: Papers II and III address the characterisation of structural changes (in stiffness and non-linear characteristics) due to environmental conditions.

Damage detection: Paper IV explores a novel model-free damage detection algorithm for bridges.

In terms of traffic monitoring, this study was limited to 1) the general characterisation of traffic, especially overloaded traffic, on a highway bridge and 2) a detailed study of the traffic on a railway bridge. Environmental variations in bridge behaviour were observed mainly in the Skidträsk Railway Bridge, and to some extent in the High Coast Bridge. The damage detection algorithm was limited to numerical experiments.

Figure 1: Schematic of the main factor influencing the structural response, and how they were examined in this work.

1.3. THESIS OUTLINE

3

1.3 Thesis Outline

This thesis is based on the work and results presented in the five appended papers. The subjects of structural health monitoring and bridge weigh-in-motion are first placed in a historical perspective as an introduction to the work. Other important contemporary contributions to these fields are also listed to set the results of the papers in context.

Recent developments in sensing technology and algorithms for bridge damage identification are reviewed in chapter 2. Other important aspects of a monitoring system, such as sensor location, communication methods and sensor failure detection, are also discussed.

Bridge weigh-in-motion is presented in chapter 3, including a history of the method and some of the latest and more important contributions to the field.

Chapter 4 discusses some of the advanced techniques used throughout this thesis for signal analysis, system identification and damage detection.

Chapter 5 briefly presents the three bridges used as case studies for the work in this thesis.

Lastly, conclusions and discussion are presented in chapter 6.

The five papers on which this work is based are provided in appendices A to E.

Paper I presents the bridge weigh-in-motion algorithm developed for the Söderström Bridge, a crucial component of the Swedish railway system. This topic is completed in chapters 3 and 4.1 with an overview of the historical and contemporary development of bridge weigh-in-motion and further information on the bridge under study. This study was published in the ICE Bridge Engineering journal.

Paper II addresses the seasonal variability in the behaviour of a ballasted railway bridge due to stiffness variations in some of the materials. The variability in the studied parameters is characterised stochastically via a Monte Carlo Markov Chain model. Possible explanations for the causes driving these changes are provided and validated by numerical simulations. This study was published by the journal Engineering Structures.

Paper III addresses the non-linear characteristics present in a ballasted railway bridge. Seasonal changes in these characteristics are also studied and parameterised stochastically. This study was published by the journal Engineering Structures.

Paper IV presents a novel damage detection approach that uses bridge weigh-in-motion data to reduce the uncertainties, due to unknown traffic load, that are inherent to damage identification. The method uses machine learning techniques to predict the deck accelerations based on previous data. Once the algorithm is trained, reductions in the accuracy of the prediction can be linked to damage. This study was submitted to Computers and Structures.

Paper V presents the monitoring system installed in the High Coast Suspension Bridge and the results obtained from that system. The topic is completed in chapter 4.2 by further information on the bridge. This study was published in the proceedings of the IAMBAS 14 conference.

1.3. THESIS OUTLINE

5

2 Structural Health Monitoring

A transportation infrastructure is vital to society, and terrestrial transport links, in the form of motorways and railways, constitute an important part of this infrastructure. Bridges are one of the critical points in every transport network, but they are expensive to build and maintain, and the consequences of their sudden failure are severe. Therefore, bridges are expected to have a higher degree of reliability, which in practice requires thorough inspection and maintenance schemes, among other measures. This requirement has led to major interest in the possibility of using structural health monitoring (SHM) in bridge engineering. Several projects supported by the European Union have already addressed the challenges and opportunities of SHM in bridge structures, providing guidelines and recommendations such as those in [2, 3].

SHM works mainly to detect, locate and quantify damage to a structure through the acquisition of data measured in situ on the bridge. SHM systems can also be used for other purposes, such as load estimation (e.g., traffic or wind), construction and repair work monitoring, and to validate design assumptions regarding the structure’s static and dynamic behaviour.

The research within SHM has been directed mainly towards the development of new sensors and new algorithms to analyse the data gathered. Bridge monitoring has been used to follow the construction stages of complex structures [4, 5], to adjust cable pressures in post-tensioned structures and for load estimation purposes [6, 7], but damage detection techniques have generally been confined to laboratory and numerical experiments. Thus, despite the advances in this area, SHM has not yet become a tool that bridge managers can use to optimise inspection and maintenance procedures.

A common classification [8] of SHM systems divides them into four classes of growing complexity depending on the damage characterisation they are capable of achieving. The first stage is the detection of damage. In this stage, the SHM system warns about the detection of a failure, without further specification on the nature of this failure. This procedure is, of course, the simplest form of SHM, and it is sufficient for many applications. The second stage consists of the spatial localisation of the detected damage, which usually requires more complex sensor networks and more advanced algorithms. In the third stage, a diagnosis of the type and extension or severity of the damage is automatically carried out by the SHM system. The fourth stage creates a prognosis for the structure’s remaining service life. Although such a prognosis would be very useful, few implementations of this fourth stage currently exist. Information on the healthy and actual (i.e., possibly damaged) state of the structure is not sufficient to create a successful estimation of remaining life because knowledge of the deterioration schemes and estimations of future loads are also required. These four stages are shown schematically in figure 2.

Chapter 2. Structural Health Monitoring

6

Figure 2. A general flow chart of a SHM system, including the 4 stages in which SHM is most commonly divided: Detection, Location, Severity and Prognosis. Modified from [9].

This chapter presents recent developments in sensor technology and SHM techniques, with a focus on bridge structures and the presentation of sensors that have been introduced recently in the field of SHM for bridges.

2.1 History of Bridge SHM (Applications)

The historical development of SHM in bridges is difficult to delimit. First, the theoretical advances in the field are most often developed as generic techniques that can be applied to different types of structures. Admittedly, most techniques are developed with a specific structural type in mind, but this does not constrain their applicability to only these structures. Some techniques are developed specifically for bridges, but are not confined to this type of structure. Therefore, a review of the theoretical advances in the SHM techniques used in bridges has arbitrary limits. It has become necessary to limit comprehension only to field deployments to obtain a clear-cut delimitation. The purpose of this section is not to present the historical development of SHM in general, so only field deployments in bridges will be listed.

Second, although some degree of automation is included in the term SHM, the exact boundary between normal inspection and what is considered SHM is not well defined and has changed over time. Although SHM could be automated in theory, most of the initial methods for performing SHM were not fully automated in practice, due primarily to hardware limitations. As the cost of computers has decreased and their power has increased, hardware limitations have been more easily overcome.

2.1. HISTORY OF BRIDGE SHM (APPLICATIONS)

7

As reported in [10], systematic inspection began in the US in 1967 after the collapse of a bridge at Point Pleasant. From then on, the use of sensors to acquire information not readily visible to the naked eye commenced in a systematic fashion. Of course, the digitisation of these methods had to wait for the informatics revolution. The first implementations of vibration-based damage detection in bridges arrived during the 80’s and, as a rule, identified modal parameters for damage identification. The studies [11] and [12] are among the earliest examples. The setup in these works was not one of continuous monitoring but was performed during a fatigue and a failure test, respectively. The monitoring used identified changes in the modal parameters as a damage indicator. In [13], a modal analysis of a composite bridge was performed, simulating the damage caused by unfastening bolts. Changes in the frequency response function were reported as detectable and quantifiable. In 1988, performed one of the earliest numerical experiments on a bridge subjected to random traffic was performed [14]. Again, the modal parameters were suggested as damage indicators. During the 90’s, several laboratory experiments were performed in beams and bridge models. In [15], continuous vibration monitoring systems is recommended, observing that the changes in modal shapes are more noticeable closer to the vicinity of the damage. In [16], it is concluded (from laboratory experiments) that frequency alone is not a reliable damage indicator, as critical damage produces frequency shifts of less than 5%. In [17], the deterioration of a railway bridge was studied, concluding that identified changes in the modal parameters provide information merely on the presence of damage, but not its location, extension or underlying cause. In [18], a pre-stressed concrete bridge is examined and it is established that for a reliable damage detection algorithm based only on frequency changes, changes on the order of 0.01 Hz must be detectable. Mode shapes, on the other hand, could be used more effectively. In [19], a failure test on a scaled bridge model is performed. Changes in the magnitude of the frequency response function (calculated from induced ground motion) were found to be good damage indicators. An experiment performed on a 3-span bridge in [20] concluded that local, non-critical damage could not be successfully detected by identifying only the lowest modal parameters, and that information regarding higher modes would be required.

In general, the first approaches to damage detection on bridges were modal based, and they compared the mode shapes and frequencies directly to observe damage. This type of damage detection is still very active, but with increasingly advanced damage-indicating features calculated from the identified modal parameters. The more commonly used features have been modal curvature [21] and the related modal strain energy [22], dynamic flexibility matrix [23] and others.

Today, structural health monitoring is a large, active research field, even when only bridge structures are considered. It is worth noting that damage detection techniques not based on modal parameters have been developed in later years [24]. Furthermore, SHM has recently been used not only for damage detection but also for continuous reliability assessments [25].

The advances of later years are beyond the scope of this chapter. A brief review of recent sensor development can be found in section 2.2 and a review of recent damage detection algorithms and applications in section 2.3.


8

2.2 Sensors Used in Bridges SHM

The SHM process starts with the measurement of relevant physical quantities in the structure. Due to the advent of modern computers and data acquisition systems, this measurement is usually achieved by a sensor that transforms the quantity to be measured (e.g., acceleration, strain, light intensity, temperature) into electrical signals that can be easily digitised and stored. Some of the most commonly used sensor types are listed in table 1. A short review of sensing technologies and their use in bridge monitoring is provided in this chapter with the goal of providing basic information for a better understanding of the current state of technology and the possibilities in bridge monitoring. A focus is placed on sensors that, without being new technologies per se, are relatively new in the field of bridge monitoring and are used preferentially in long-term monitoring as opposed to temporary instrumentation.

Table 1: Typical sensors used in structural health monitoring (from http://www.sustainablebridges.net/)

Physical quantity Sensor

Displacement Linear variable differential transformer (LVDT) Long gauge fibre optics Optical Laser

Acceleration Piezoelectric accelerometer Capacitive accelerometer Force balanced accelerometer MEMS

Strain Electrical resistance strain gauge Vibrating wire sensor Bragg grating fibre optics Long gauge fibre optics (interferometry)

Force Electrical resistance load cell Piezoelectric load cell

Temperature Electrical resistance thermometer Thermocouple Thermistor Fibre optics based sensor

2.2.1 Cameras

Cameras have been used to measure deflection in bridges under thermal loading, dead and traffic loads, crack lengths and widths, and to monitor corrosive damage [26-28].

Obtaining the geometrical properties of objects is usually referred to as photogrammetry, which is a non-destructive, remote sensing technology that can be rapidly deployed in different structures and without elevated costs. With the general digitisation of cameras and the inexpensive availability of computer power, photogrammetry has become a more extensively used practice in many fields. The possibility of directly measuring displacements, as opposed to strains or acceleration, is a very attractive characteristic of photogrammetry.

2.2. SENSORS USED IN BRIDGES SHM

9

Even though most applications are static, some implementations of dynamic, real-time, vision-based measurement in bridge structures exist. The analysis of visual information gathered by camera arrays can be relatively easily done with commercial software developed specifically for that purpose. Algorithms for target recognition and motion extraction have been successfully used in bridge monitoring for dynamic measurements.

In [29], an off-the-shelf video recorder was used in conjunction with a telescopic arrangement to take a motion picture of a target consisting of 4 white dots forming a square on a black background. The motion of this target could be tracked on-line by digital image processing techniques. This optical monitoring system is schematically described in figure 3. The arrangement was validated by studying the dynamic behaviour of a 4-storey structure and comparing the results with those obtained by more traditional fibre optic sensors. Both methods registered the same displacement with up to a 3% error. A field test was also performed on a 4-span, open steel box composite bridge. Loaded trucks with total masses of 30 and 40 tons were run across the bridge at speeds of 3, 20 and 40 km/h. The camera was placed 20 m from the target and its sampling frequency was 30 Hz, which was sufficient for accurate representation of the bridge dynamics because most of the energy content was under the 3 Hz threshold. The results were compared with those obtained by a laser vibrometer, with satisfactory results in both the time and frequency domains. This method allows for a relative inexpensive measurement of displacement with high space resolution and good frequency resolution. One of the main drawbacks is the necessity for clear visibility and high sensitivity to even small camera vibrations because the effects will be magnified by the large distance to the target.

In [30], an algorithm was developed to use changes in the structural characteristics, detected by a high resolution camera, to diagnose damage in the form of stiffness reduction. The method is depicted in figure 4. The sensing and data analysing array was capable of accurately detecting and locating damage corresponding to a stiffness reduction of 3% under laboratory conditions. To this end, the digitised visual data (monochromic light) was polynomial-fitted so that sub-pixel accuracy could be achieved. Some mathematically relevant points (such as inflection points and local maxima/minima) were detected in both the damaged and healthy conditions. From this information, physically meaningful quantities were derived (displacement, slope, curvature) and changes in those quantities were used to calculate a damage index at each location.

Figure 3. Illustration of remote structural deflection monitoring using an off-the-shelf video recorder. Modified from [29].


10

Figure 4. Optical-based displacement monitoring. A) Picture of the experimental set-up taken with a high-definition camera. B) Close-up of the highlighted loaded area. C) Plot of light intensity for the pixels outlined in B. The polynomial fitting (black line) of this discrete data point allows for sub-pixel accuracy. Modified from [29].

In [31], an algorithm to inspect surfaces in search of cracks is developed. The installation of cameras can reduce the risks and elevated cost associated with the inspection of surfaces that are difficult to access, such as bridge soffits. In a field experiment, these researchers installed the sensing system on a crane mounted on a truck to survey different bridges. The method was not found to be reliable enough to replace human inspection, but it could be used in places with limited accessibility, or to inspect critical points with a high cracking risk or with dangerous cracks already present.

2.2.2 Fibre Optics

The most commonly used fibre optic sensor (FOS) for measuring strain is the interferometric FOS. In this sensor, the light is divided into two beams, one sent through the measuring strand and the other through a passive reference strand. When the beams are recombined, the relative phase differences can be measured and associated with a given physical value (most commonly strain or displacement) [32].

FOS’s are immune to electrical disturbance, and have a high resistance to corrosion, long-term measurement stability and very high measurement accuracy. Moreover, they are easier to embed in different materials than other sensor types. In addition, they can measure different physical quantities at the same time (typically temperature and strain).

Different fibre optic techniques measure strain in different fashions. The intensity, phase shift and wave length of the reflected beam can be measured and translated into relevant structural


11

parameters (mainly strain or relative displacement, but relative velocity, temperature and pressure can also be measured with different configurations). The cost implicated in these different methods varies significantly. The most advanced techniques, which use demodulation, have a cost that is prohibitive for most SHM applications, but inexpensive methods of interrogation are also available.

FOSs have been developed to measure a number of physical and chemical variables besides strain. Humidity, corrosion, pH and chloride sensors have also been studied in the past [33].

Distributed Fibre Optics

FOSs can measure distributed temperature and strains along their entire length by virtue of the Brillouin scattering effect. When a pulse of light interacts with thermally excited photons within the fibre or with changes in the refractive index due to strain, a frequency shifted reflection of the pulse propagates backwards in the optic fibre. By measuring the amount of frequency shifting, the strains can be calculated. If the time it takes the light pulse to travel forwards and backwards in the fibre is considered, the exact point where the reflection takes place can be calculated. The frequency shift carries the information about the strain and temperature at a given point. Temperatures and strains cause different amounts of light to be backscattered and different frequency shifts for different input frequencies (colours). Thus, by using input light of different frequencies, the effects of the temperature and strains can be separated. Commercially available systems have spatial and thermal resolutions of approximately 1 m and 1 °C, respectively, which is satisfactory for many bridge applications. This type of sensor can monitor extremely long distances on the order of kilometres [34].

In [35], a distributed FOS was used to monitor the distributed stresses in the cables of post-tensioned concrete beams. The sensor used was a Brillouin optical time domain reflectometer (BOTDR), which measures strains and temperatures in arbitrary regions of an optic fibre strand, as shown in figure 5. In the laboratory test, two beams were prepared, one of which was strengthened with two external post-tensioning cables and the other of which was strengthened by a single bonded cable. The different beam configurations were loaded with stepwise increasing loads and the tension in the cables controlled by both strain gauges and BOTDR. Both systems agreed to within 2.7%. From this study, it could be concluded that cable tension suffers considerable local variations that can go unnoticed by a traditional strain gauge detection system, which only measures at a given point, in contrast to the BOTDR, which gives reliable information about the strains along the entire length of an optic fibre.

This type of sensor has also been used successfully for validating strengthening methods, such as in [36], where the distributed character of this type of sensor was especially useful for controlling the adhesion of strengthening elements along its length.

Bragg Grating Sensors

Fibre Bragg grating (FBG) sensors are a successful and relatively new type of FOS. In these sensors, a special fibre optic is modified by creating periodic variations (called gratings) in its refractive index. Part of the light that passes each of the modified zones is reflected back. The periodicity of the modified zones will cause a certain wavelength to be reflected in phase and thus amplified, as depicted in figure 6. A structural change in this period, due to strain or temperature, can then be measured as a change in the reflected wavelength. A wavelength, in contrast to phase-shift, is an absolute parameter and is therefore less affected by imperfections


12

in the power input or beam path, making the FBG a more reliable sensor. These sensors are also, in general, of lower cost compared to other FOS types because the interrogation methods do not require interferometry or demodulation. Grating with different periods can be introduced in the same fibre, allowing for the measurement of multiple points [32, 36].

Figure 5. Principle behind the BOTDR in which segments of the fibre with different strains scatter light at different frequencies. From [35].

Figure 6. Scheme of a FBG, and its effect on the spectra of transmitted and reflected light. Modified from [32].


13

2.2.3 Electrochemical (Corrosion)

The corrosion of steel rebar embedded in concrete is one of the main causes of damage and failure in reinforced concrete structures. In the corrosion process, metallic iron transforms into iron hydroxide, which has a larger volume. The pressure exerted creates cracking and spalling of the concrete around the rebar. Corrosion thus leads to a reduced reinforcement cross-section, loss of contact between the rebar and concrete and the formation of new cracks that allow more corrosive agents to penetrate the concrete, accelerating the entire process.

The most common type of damage in Swedish railway bridges reported by the former Swedish Rail Administration, now the Swedish Transport Administration (Trafikverket), is corrosion (16% of cases), followed by spalling (8% of cases). As can be seen, corrosion occurs at double the frequency of any other damage type. This ratio could be even higher for highway bridges that require salting during winter. At least 700 of the Swedish Transport Administration’s close to 4000 railway bridges have reported corrosion damage in load-bearing elements. In 35 of these cases, the damage has been classified as “condition class 3,” meaning that it should be attended to immediately.

Corrosion reveals itself in a number of ways, such as though cracks or changes in the electrical or chemical properties of both the steel and the concrete. Thus, many different approaches to its detection can be used. Several different sensors have been used successfully, and novel sensors are being developed on a regular basis.

In [37], a novel corrosion sensor based on the measurement of MnO2 is presented. It is showed to have comparable accuracy and better long term stability when compared with traditional SCE electrodes. In [38], a galvanic current sensor is developed. Rather than actual corrosion, this type of sensors measures the corrosion rate because the process of corrosion causes current to flow within the rebar. The results are more difficult to interpret, but can give useful extra information when combined with other corrosion monitoring methods. In [39], a corrosion sensor, is developed. It is based on the electric response of rebar to an applied galvanic current pulse. The decay rate of the potential in the steel, after a current pulse was applied, was found to be a good indicator of the corrosion level in the rebar.

Figure 7. Probability densities for different local damage events are input (this example shows the probability of corrosion initiation in a structural member, left-hand plot), and from these densities, global probability densities are calculated (this example shows the total reinforcement area in the relevant structural member, right-hand plot). From [40].


14

Corrosion is a very local phenomenon, and the critical parts in a bridge with respect to corrosion are not always obvious. The amount of corrosion is in practice impossible to measure at every point in space and time. Therefore, models used to estimate the probability of corrosion in unmonitored areas, from the limited spatial information obtained in the sensor-equipped areas, are important to a structure’s safety assessment of. Also, models of the temporal evolution of corrosion from a given measured actual state could have a large implication in, for example, the estimation of the structure’s remaining service life. The probabilistic model presented in [40] estimated the reinforcement loss due to corrosion in concrete structures based on the temporal extrapolation of data obtained from embedded corrosion sensors, as shown in figure 7. Both the optimal placement and the type of sensor were discussed, and the model was used to calculate a more accurate partial coefficient for the safety of reinforced concrete structures.

2.2.4 Laser Doppler Vibrometer

Laser Doppler vibrometers (LDV) are a sensing technique based on the frequency shift produced in a light beam when it is reflected on a surface moving relative to the emitter. They can measure speed, and displacement to a resolution of less than one hundredth of a millimetre, with a sampling frequency in the MHz range. LDVs are remote, meaning that no physical attachment to the structure being measured is required. They also allow for measurement in parts of the structure that are difficult to access. Measurements can be taken for up to 30 metres with no significant loss in accuracy. Many commercially available LDVs can measure different points on a surface within 10 milliseconds of each other, allowing for great spatial resolution [41]. The main drawback to the use of LDVs in bridge monitoring is their cost, which renders them impractical for permanent installation in the structure to be monitored. In general, LDVs are dependent on good visibility levels, which are not always guaranteed in outdoor conditions and further reduce their usefulness in long-term SHM.

2.2.5 Accelerometers

Accelerometers are, together with strain gauges, the most commonly used sensors in SHM. The usual configuration consists of a small mass resting on a sensing element (e.g., made of piezoelectric material). As the frame to which the accelerometer is attached accelerates, the inertia of the mass produces deformations in the piezoelectric base. This deformation induces electrical currents that can be measured and interpreted back to acceleration. Extensive literature reviews on the subject of acceleration-based SHM can be found [42, 43].

2.2.6 Strain & Relative Displacement Sensors

Resistance Strain Gauge

Since the invention of strain gauges in 1938, these instruments have been used extensively in civil engineering. Strain gauges are simple, reliable, linear in their behaviour and extremely inexpensive. A piece of conducting material becomes larger when subjected to tension and, because of Poisson’s contraction, also thinner. Conversely, if compressed, it becomes shorter and thicker. Both effects result in changes in electric resistance that can be easily measured and translated back into strain. Strain gauges can be installed directly on the surface of a


15

bridge element or embedded inside the element for measuring internal strain. They can be preassembled on normal reinforcement bars as so-called sister bars and placed alongside the main reinforcement in a concrete element.

Vibrating Wire Sensors

Vibrating wire sensors work much like a guitar string. An elongation of the wire will increase the normal forces in it, changing its eigenfrequency. In this manner, the relative displacement of the wire’s ends can be inferred by measuring its eigenfrequency.

Linear Variable Differential Transformer (LVDT)

Current is transmitted through an LVDT sensor, via the transformer core bar, from a primary winding to two secondary windings. These secondary windings are placed coaxially with the primary winding, on each side of it. The transformer core moves freely along the axis of the 3 coils, connecting only partially to the primary and secondary windings. The difference in the voltage transmitted to each of the secondary windings depends on the position of the core bar. LVDTs have a long life, high resolution and high signal-to-noise ratio.

2.2.7 Temperature Sensors

Temperature is an important variable in SHM. In addition to the weight of the bridge itself and that of vehicles crossing it, temperature is the most crucial load affecting a bridge’s behaviour. Temperature gradients induced by uneven exposure to sunlight and other effects can have non-trivial consequences in the measured response of a bridge to a given load [8]. Therefore, methods to estimate or filter out this effect have been studied since the beginning of SHM development. It has been noted that temperature effects can produce frequency changes in excess of 14%, which is comparable to the frequency changes induced by severe damage to a structure. In comparison to other physical quantities, such as acceleration or strain, these changes are very easy and inexpensive to measure. Unfortunately, many proposed SHM methods have discussed the effects of temperature only superficially.

2.2.8 Acoustic Emissions

Acoustic emission sensors are designed to capture the sound waves that spread through a material when certain events take place. In essence, they are a modified microphone applied to the surface of (or embedded in) the structure being monitored. Distinguishing the type of event causing the recorded acoustic emission, its location and the maintenance of an “event count” to estimate the accumulated damage are the most important components of acoustic emission SHM. Therefore, this technique is addressed under the Methods section (see section 2.3), although it makes use of its own set of sensors.


16

2.3 Algorithms used in SHM of bridges

This section presents recent developments in SHM and damage detection techniques. The paper focuses on bridge structures and the presentation of sensors and algorithms that have been introduced recently in the field of SHM for bridges, although they are general algorithms not limited to bridge structures.

2.3.1 Data analysis & evaluation

Many interesting structural properties, such as damping, non-linearity in response, modal shapes and frequencies, are not directly measurable and must be inferred from other measurable data. Although these variables cannot be measured directly, they possess (at least theoretically) clear definitions that allow for the design of relatively straightforward algorithms to quantify them. Of course, more complex and advanced algorithms may achieve a higher degree of robustness or accuracy, but, in general, simple methods exist to obtain coarse approximations of these properties. Damage identification is not so simple because it can be difficult to describe what is meant by damage, and even more difficult to put it in mathematical terms. Damage detection methods are very specific, not only to a structure, but even to the type of damage that is being monitored. It is therefore desirable to identify a “damage index”, or a function that will relate the actual health state of the structure to a non-negative number. Li et al. [44] present a novel method for damage detection based on the increased fractal dimension of the identified eigenmodes. The method is tested numerically and in a laboratory experiment, and is capable of detecting and locating damage in a simply supported beam, even when multiple instances of damage are introduced. A serious limitation of the method is that the estimation of the eigenmodes must have a very high spatial resolution. Carey et al. [45] repurpose a moving force identification algorithm as a damage detection tool. The idea behind the proposed method is that, if damage is present, a moving force identification (which works under the assumption of a healthy bridge) will return load estimations that differ significantly from the estimation under healthy conditions. In particular, the identified forces (from damaged scenarios) seem to have a linear trend not present in the healthy case. The method is tested in a numerical experiment. Liu et al. [46] use ambient vibration data from the Xing Nan Bridge before and after major repair work to validate different approaches to damage detection. Among the methods tested, the Hilbert-Huang transform with empirical mode decomposition is shown to be capable of discerning the signals from before and after the repairs. Ensemble empirical mode decomposition is also tested and shown to perform better than empirical mode decomposition. Kim et al. [47] present a hybrid global/local, 3-stage method for damage detection and quantification. In the first stage, a vibration-based damage detection algorithm uses frequency response functions to identify global changes in behaviour. In the second stage, electro-mechanical actuators/sensors are used to measure changes in the local impedance and classify damage into tendon or girder damage. The third stage uses modal parameters to estimate the

2.3. ALGORITHMS USED IN SHM OF BRIDGES

17

severity of the damage. The method is tested with a stressed concrete beam in a laboratory experiment. Rodrigues et al. [48] present a real-life deployment of a long-term monitoring system that uses fibre Bragg grating. The monitoring system comprises strain and vertical displacement gauges, and the deployment complements more traditional structural health monitoring in the same bridge. The purpose of the FBG system is to characterise the structure’s response to traffic and environmental conditions.

Acoustic Emissions

Structural defects, such as fatigue cracks, emit ultrasonic stress waves that are easy to convert into electrical signals with the appropriate sensors. These signals can then be used to identify and locate the defect. This identification is a different approach compared to more traditional damage detection because it detects the nature and active moment when the damaging process occurs, rather than the damaged status of the structure. This quality is also one of the method’s major disadvantages because there is no simple method to relate the damage processes taking place to an accumulated damaged status. Reference [49] discusses the potential of acoustic emission (AE) sensing for bridge monitoring, giving a general overview of the method’s advantages, disadvantages and possibilities. Among the advantages listed are the following: it is completely passive, it detects the exact moment when damage occurs, the sources of damage do not need to be known with accuracy, and it can cover large areas with few sensors. The method has been proven capable of tracking fatigue crack information with accuracy in both laboratory and field experiments. Crack propagation monitoring is one of the most common uses of AE in concrete and steel structures [50, 51]. One of the challenges of acoustic emission monitoring is in locating the source of the acoustic waves. As the complexity of the structure under study grows, the location of the sources becomes increasingly difficult. In [52], a new method was proposed for complex metallic structures that greatly increased the accuracy of the source identification in non-trivial geometric shapes. The method requires considerable amount of training (although less than other proposed methods), which could complicate its implementation in bridge monitoring. More relevant to bridge engineering is the work developed in [53], where Rayleigh waves were used to determine the source of acoustic emissions in large plate-like concrete structures in a laboratory experiment (see figure 8). The use of Rayleigh waves instead of P-waves gave the system a larger range and allowed it to detect and estimate the source of emissions with practically undetectable P-waves. In [53], AE monitoring was implemented in a number of concrete structures and laboratory experiments, including a bridge. The damage detection approach was simple and did not locate the source of the emission, but simply estimated the amount of energy released as an indirect measurement of the damage. A study of the effect of the studied body’s size was performed, leading to some size-independent parameters being used to define damage levels. In [54], a similar experiment was carried out in steel structural members. The crack propagation was compared with the count rate of acoustic emissions above a certain threshold to investigate the relationship between the two. Material plasticisation, crack closure and other phenomena that also produce acoustic emissions tended to complicate the relation between crack propagation and the acoustic emission rate. Nair and Cai [55] presented a recent literature review on the acoustic emissions used in bridge structural health monitoring.


18

Figure 8. Concrete structure monitored with the aid of acoustic sensors. In this experiment, Rayleigh waves were used to detect the source of acoustic emissions from cracks and other events. From [53].

Modal Analysis

Vibration in structures with mainly linear behaviour can be decomposed into a number of modes. The shapes of these modes, and their corresponding frequencies, are a function of the mass distribution and the structure’s stiffness. Because mass often remains unchanged, even under damaged conditions, changes in mode shapes and their associated frequencies are a good indicator of stiffness changes, and are usually caused by damage. There are several methods for the identification of modal parameters, and these methods are of very different complexity and accuracy, but the field is a very widespread and well-understood branch of structural engineering. The global nature of modal parameters allows for the detection of damage even if the specific location of the damage is not instrumented, keeping the costs of monitoring systems low. At the same time, this very fact makes the localisation of damage a difficult task. In [56], changes in frequency are used as an initial detection of damage and a double criteria method to localise the damage is used. The double criteria damage index was based on the modal strain energy and modal flexibility matrix change. These two criteria were derived from the structure’s modal parameters. The method was applied to numerical simulations in both beam- and plate-like structures. Both proposed criteria were found to work well in single-damage scenarios, but a considerable enhancement of the localisation capabilities was noted in multi-damage scenarios when the results obtained from the different methods were combined. In [57], a method to experimentally determine a structure’s flexibility matrix is proposed. It was combined with virtual load, could be used to detect changes in the stiffness in small regions, usually indicating damage. The quasi-static flexibility matrix used in this approach


19

can be directly calculated from the modal parameters, with no need for controlled test loading. The method is general and could, in principle, be used in any kind of structure, but the authors concentrated on beam-like structures. Simulations and experiments were used to validate the method, which was found to be accurate compared to an updated FE model and direct stiffness calculation approaches in simple beams. Nonetheless, it was difficult to apply the method to more complex structures with varying mass distributions or significant stiffness changes in different sections. Particular care must be taken with the number of modes necessary to obtain accurate results because this number was shown to vary with the type and extent of damage. A vibration-based damage detection and sizing method was developed in [58]. The method considers frequency changes due to temperature effects to avoid false positives, and is based on the creation of statistical control charts that describe the variation of the eigenfrequencies with temperature. It can determine whether an introduced frequency change matches the pattern produced by a temperature variation, or if it should be considered a novelty and therefore an indication of damage. The accuracy of the algorithm is greatly enhanced when the actual temperature of the measured structure is known, but it is not completely necessary for damage detection and a fairly accurate localisation of the damage. In [59], a novel approach to detecting changes in the linear dynamic behaviour and eigenmodes of linear lightly-damped structures is presented. This approach does away with the need for eigenfrequency/eigenvector analysis. Instead, statistical properties of the provided signals were used with the aid of modal power shapes. These features are similar to modal shapes (see figure 9), but are based on signal power spectral densities rather than modal parameter extraction. The technique worked satisfactorily for structures with eigenfrequencies distant from each other, but was somewhat problematic to apply in complex structures that require a high density of sensors, or in structures with different modes of similar frequency, because it numerically integrates the frequency content of each modal peak. Using the power spectral density, a measure of the power mode shape curvature and power flexibility is calculated and used as a damage index to identify, locate and somewhat estimate the presence of damage in a structural element. The capability to detect and locate damage was proven in a number of numerical simulations and experimental arrangements. The method successfully detected and located minor damage, even in the presence of high noise levels.

Figure 9. Modal shapes from a slender beam obtained by traditional modal analysis (left plot) and by modal power (right plot), as proposed in [59].


20

Magalhaes et al. [60] offer an example of a real-life deployment of a long-term monitoring system used for damage detection. They extracted the frequencies of the lowest eigenmodes and evaluated them in a regression model that accounted for the effect of external environmental factors such as temperature and traffic intensity. Once these effects were isolated, the eigenfrequencies were found to lie within a narrow band for all temperatures and all traffic conditions. Deviations from this trend could be interpreted as damage. A numerical experiment showed that stiffness reductions of 10% could be detected with the proposed method. Radzienski et al. [61] presented a novel damage detection technique based on identified modal parameters. The modal curvature, COMAC, strain energy, modified Laplace operator, fractal dimension, wavelet transform, potential strain energy and frequency shift methods were also explored and compared. A correction factor was also provided for many of the presented methods to improve their results. The proposed methods were tested in a laboratory experiment on a cantilever beam, and a hybrid method that combined the other presented methods was developed, tested and recommended.

Statistical Pattern Recognition

Vibration-based SHM is essentially a statistical pattern recognition problem. Data on the structure’s behaviour are collected and analysed with statistical tools that permit the detection and classification of changes. The principle is to detect changes in the behaviour caused by damage at an early stage. However, behavioural changes due to damage are usually of lesser magnitude than those due to normal loading and environmental effects, except in the case of very severe damage. Therefore, an accurate knowledge of the healthy (undamaged) structure’s behaviour is needed to successfully recognise damage-induced changes. Common methods of acquiring this knowledge are through finite element modelling (FEM) and neural network (NN) training. A possible, and very common, approach to SHM involves measuring the dynamic behaviour of a structure and then comparing it to a simulated behaviour obtained from a numerical model. This approach becomes prohibitive, even for very simple structures, due to the uncertainties that are always present in real-life structures and the difficulty of determining these uncertainties a priori to introduce them into the model. Another approach that has attracted increasing interest from researchers is called statistical novelty detection. Various statistical characteristics of the structure’s dynamic response are studied, and algorithms that detect changes in these characteristics are developed. Common to these methods is a “learning” phase, in which data from the healthy structure must be provided to create a comparative framework against which future, possibly damaged, data can be analysed. Genetic algorithms and NNs are used as a computationally efficient way of extracting the significant parameters from the data and “teaching” the algorithm what a healthy signal should look like without requiring comprehensive previous knowledge of the structure under study. In [62], two statistical novelty measurement methodologies are implemented in experimental arrangements to validate them. First, free decay responses for the studied structures were generated from ambient excitation tests using random decrement. Next, the auto regressive


21

model that best fit the data was identified. An auto regressive model estimates the value of a variable at a given time as a linear combination of a given number of previous values. Differences in the coefficients of the auto regressive model are used as a measure of the novelty in the signal, and therefore of the damage. This approach is then tested in experimental arrangements for beams and grid-like structures. The method was found to be capable of detecting and, to a certain extent, locating the damage introduced on most occasions. It was noted, however, that some damage configurations went undetected, and that a more robust determination method for the threshold was needed. In statistical pattern recognition approaches to SHM, a damage index is usually derived from statistically relevant data. These data map an entire set of possible structural states into one single real number that, when exceeding a given boundary, will be classified as a damaged state. Correct boundaries for this index are critical to the success of SHM schemes. A parameter estimation technique was developed in [63] that could be automatically applied to measured data. This technique identifies the underlying probability distribution of the extreme values of an observed variable without a priori assumptions about its nature. The method allows for a more objective decision boundary without subjective user intervention. Mustafa and Necati [64] presented a novel time-series approach to damage detection. The proposed method used an auto-regressive with exogenous output (ARX) model with sensor clustering. In this model, an algorithm chooses which sensors to cluster, after which the outputs on the other sensors in a cluster are used as exogenous input for each sensor ARX model. The proposed model was tested on a numerical benchmark and in a laboratory experiment on a beam structure. The following subsection describes important pattern recognition methods.

Genetic Algorithms

Genetic algorithms are a computational technique used for solving optimisation problems. In a genetic algorithm, an initial set of parent solutions are ranked by fitness or quality (i.e., how well they solve the problem at hand). Based on their fitness, a stochastic selection and recombination of parent solutions occur to produce a new generation of solutions, much like the genome of a living being recombining in reproduction. Among the resulting solutions there will be those with better fitness than the parent solutions and those with worse fitness. Those with higher fitness are more likely to be chosen for the creation of a new generation, so the process will tend towards the optimal solution. Genetic algorithms have major advantages when solving certain types of optimisation problems, especially pattern recognition problems. In [65], the vector of eigenfrequencies and its changes to detect and locate damage is used. Using a model of the structure, the rate of change in the different eigenfrequencies could be calculated as different damage scenarios were introduced. This process allowed for the possibility of detecting multiple damage scenarios, and even estimating the level of damage depending on the changes to the eigenfrequency vector. Given the number of possible damage level and damage location combinations, a genetic algorithm is the only feasible way of finding the right combination of damage scenarios that will result in a measured change in the eigenfrequency vector. This genetic approach was compared with others, such as least squares and frequency-error, and shown to be overwhelmingly superior. In the same article, the stacked vector of mode shapes and its correlation was used as a damage indicator in a similar fashion. This new method accurately detected and located areas of multiple damage


22

corresponding to a stiffness reduction of 10%, even in the presence of moderate noise levels (2% standard deviation). Stacking enough mode shape vectors was found to be critical to the method’s accuracy. Damage scenarios were very accurately located in 28 eigenmodes, but were poorly located in 20 eigenmodes. Jafarkhani and Masri [66] implemented an evolutionary strategy to update a finite element model and characterise the state of a monitored bridge. If novelty is detected, an ARMA model triggers a subspace method for system identification used on vibration data. A model is then updated following the proposed evolutionary strategy to match the modal (i.e., shapes and frequencies) characteristics of the structure. The method was tested in a laboratory experiment in which damage was introduced by forced vibration. Nobahari and Seyedpoor [67] proposed an optimisation procedure based on a modified genetic algorithm to update a finite element model and interpret changes in the eigenfrequencies as stiffness reductions in certain parts of a structure. The genetic algorithm was modified, adding the assumption that the damaged elements were few in number (normally the case), which could be used to lead the mutations towards faster convergence. The method was tested numerically in beam and truss structures.

Meruane and Heylen [68] presented a hybrid algorithm that combined local optimisation with a global genetic algorithm to update a model for damage detection. A number of different objective functions based on frequency shift, modal difference, MAC values, strain energy and modal flexibility were tested and compared, and the proposed method was tested in a laboratory experiment on a 3D truss structure.

Neural Networks

Pattern recognition is a task at which computers do not excel. Today’s most promising approach in this respect is the artificial neural network (ANN). Originally developed to mimic the behaviour of a biological neural system, ANNs are proven to have good pattern generalisation capabilities. ANNs must be fed a selection of physical features to search for patterns and have a training algorithm specified that can be used to teach the network to find these patterns. Both tasks are non-trivial and often require expert judgment and experience with ANNs. In [69], a mathematically rigorous algorithm was provided to design the main parameters of an ANN (namely the activation function and the number of neurons in the hidden layer) for a given number of input and output neurons and with consideration given to the quantity of training data available. This approach avoids the need for subjective judgments, while providing ANNs that are, in a sense, optimal for the task at hand. In [70], a sequential scheme to detect and locate damage in beam structures is used. First, a neural network was trained to detect novelties in acceleration signals from cross-correlations calculated by sensors at different locations. The training phase used was the so-called supervised training, in which signals from healthy and damaged conditions are provided to the training algorithm. These signals are obtained from a numerical model of the monitored structure. In a second stage, the modal parameters of the structure were extracted. Modal shapes were used in this phase because they are less sensitive to temperature and other environmental changes. Another neural network was trained to recognise changes in the modal shape and locate damage from them.


23

Figure 10. Flow diagram of the sequential damage identification using ANN. From [70]. The algorithm was tested on numerical simulations and an experimental set-up, and was proved capable of locating and estimating multiple damage locations corresponding to very small stiffness reductions. A schematic of the monitoring system is depicted in figure 10. In [71], a fuzzy pattern recognition algorithm is used to study the novelty of signals, which could be interpreted as damage introduced in the structure. The algorithm used wavelet transforms to separate the signal into a number of components that could then be compared with healthy signals to detect changes in the structure’s mechanics. When using wavelet transforms, a high-pass filter separates a signal into “details” (higher frequency content) and a “coarser approximation” (lower frequency content). This approximation can also be separated using a rescaled version of the original filter. By repeating the (rescaled) filtering, the signal can be separated into any number of detail levels in a process illustrated in figure 11. This method gives a representation of the signal’s frequency content at different times, avoiding many of the difficulties associated with the Fourier transform. In [71], an unsupervised trained neural network was used to detect changes in the wavelet coefficients of the signals. A damage index was derived from the difference between the healthy and actual wavelet coefficients. The damage index’s fuzzy membership to a different damage level set was then calculated, and the functions that describe the membership in the different fuzzy set were changed with each new measurement by Bayesian updating. This updating process allowed for the contribution of both expert judgment and new data.


24

Ko et al. [72] described the instrumentation of the Ting Kau Bridge and the associated data gathering system. They used data gathered by the system to train an artificial neural network to predict the measured eigenfrequencies as a function of 20 different temperature inputs. They also studied the effect of wind forces on the eigenfrequencies. A number of different strategies for damage and novelty detection were proposed (auto-associative ANN, probabilistic neural networks, etc.) and tested in a numerical experiment. Jiang et al. [73] made use of modal parameters to input into a fuzzy neural network that returned a rough structural assessment. A number of these assessments were then combined in a second stage and fused into a damage indicator of greater reliability. The modal information used in the paper consisted of the frequency change ratios together with mode change ratios and a damage signature index. The neural network was trained in a supervised manner and the approach tested on a numerical example.

Information Theory

Information theory is a branch of applied mathematics and signal analysis that was developed as electronic communications commenced. To a certain degree, the interaction between different parts in a structure can be seen as information exchanges because the state of one part will eventually affect the other, and vice versa. The way these parts communicate is defined by several structural parameters, particularly stiffness, so a change in this parameter can be detected as a change in the information shared by the structural parts.

Figure 11. Flow chart describing the algorithm used to compare the wavelet coefficients of a measured signal with that simulated by an ANN. At each step, the signal is divided into an approximation part (SA) and a detail part (SD) that is subsequently divided again. The detail parts obtained from the measurement are compared to those from the ANN to estimate damage location and severity. From [71].

2.4. OTHER ASPECTS

25

In [74], information theory principles are used to calculate time-delayed mutual information and transfer entropy as a measure of the coupling between structural components. Damage that introduces nonlinearities in this coupling could be detected and the degree of nonlinearity quantified and used as a measure of the degree of damage. One main problem in vibration-based SHM is the fact that normal ambient fluctuations can cause changes in the properties of the structure, which can be even larger than that caused by significant damage. Temperature alone, as reported in [7], has been shown to change eigenfrequencies in bridges by as much as 5% over a 24-hour period. These environmental changes do not affect the linearity of the behaviour. Therefore, the detection of introduced nonlinearities can successfully discriminate between normal ambient fluctuations and damage-induced changes. In the experimental arrangement, a linearly increasing change in the stiffness was introduced to simulate a temperature gradient. This change would have been falsely detected as damage by modal analysis, but because no nonlinearities arose from the change, the method could still detect damage that produced changes in the modal variables lower than the temperature gradient. The proposed method used Gaussian random excitation, which has both advantages and drawbacks. In reality, the excitation is seldom completely random, and to presume it is may lead to bias. At the same time, ambient excitation is usually accurately approximated by Gaussian noise, so no further information about the input forces is required, thus avoiding the cost of a fully controlled test load. In [75], the major detection and localisation potential of a transfer entropy modification is shown. The transfer entropy was calculated using double time-delay, leading to a more sensitive but more computationally consuming method that did not assume the dynamic of the system to be time-invariant. With the aid of a numerical simulation and experimental arrangement, this modified version of the transfer entropy was shown to be at least as sensitive, if not more, to structural changes as the more traditional modal analysis (in both the time and frequency domains). Furthermore, the modified transfer entropy had a higher capability to locate the damage than the other damage measures studied. Fan and Qiao [76] presented a recent literature review on vibration-based damage identification.

2.4 Other Aspects

Structural health monitoring includes many other aspects beside sensor choice and a damage detection algorithm. Among these aspects, sensor placement, sensor failure detection and data communication paradigms have been treated in the literature. Some of the latest contributions to these fields are described in this section.

2.4.1 Data Communication

Data communication occurs in at least two stages of the most common SHM implementations: first from the sensors to a central storage and analysis assembly, via a data acquisition system, and then from the central assembly to the end user, usually the bridge


26

owner. Technically, the first of these stages contains more challenges because data storage and analysis can be placed in a site that facilitates communication with the outside, while the sensor placement is constrained by the structure itself and by the variables chosen for monitoring. The most common solution to the sensor-central computer communication dilemma is to hardwire each sensor to the central storage unit. Most of the time, power supply needs will require a cable to be drawn to each sensor in any case, so there is no way to avoid this problem, which can be very difficult when large structures are being monitored. Research on wireless communications and autonomous sensor networks has recently been carried out, and a number of wireless monitoring systems have been reported in the literature, but extending the monitoring period, even to months, can be challenging due to the stringent hardware limitations imposed on wireless systems [77].

Wireless sensor networks provide a variety of new possibilities for monitoring structures. In many cases, it is unpractical or even impossible to connect the different components of a sensor network by wire. This could be the case if the components are situated at long distances or if they are located in difficult-to-access places. With today’s communication technologies, wireless data transfer is relatively inexpensive and very reliable. Although limitations in the amount of data per unit of time are more severe than for wired communications, commercially available communication systems will suffice for almost any static measurement purposes, and also for many dynamic measurements.

A main problem with wireless sensor networks is their power supply. Today’s systems are almost always powered by batteries that need periodic replacement, which is expensive and demands that the sensors are easily accessible, limiting the possibilities for sensor placement. Transmitting wirelessly costs energy, so reducing the amount of data to be transmitted can save energy. Systems have been implemented in which the data are partially analysed before being transmitted, reducing the amount of data that needs to be transmitted. This approach poses problems of its own because the sensors must have enough computational power to carry out the calculations by themselves, which is also energy consuming and increases the cost of the sensors. Hybrid arrangements, in which sensors are hardwired to a temporary central computer that carries out the first calculations and transmits wirelessly for their final manipulation and storage, have also been studied.

In [78], a complete wireless monitoring system is presented. The sensors used were capacitance-based, as opposed to the conventional voltage-based sensors, meaning that they measured in a completely passive manner. Furthermore, the displacement sensors were designed to mechanically store the peak displacement values, with no need for a power supply. To retrieve the data, the sensors are connected to a communication unit that extracts the value stored in each sensor and then sends it wirelessly to a mobile station. The power required for this communication process is wirelessly supplied to the communication unit from the mobile station via a high-energy radio signal. The system is depicted in figure 12. The energy from this signal is absorbed and stored in a capacitor. When enough energy is provided, the process of reading each sensor and sending the measurements is triggered, completing the cycle. This method was implemented in the Alamosa Canyon Bridge in New Mexico. The mobile unit that provided the power and received the signal from the communication unit in the bridge was mounted on a car that could then retrieve the values stored by the system by simply passing over the bridge.

A technique called synchronised switch harvesting on inductor (SSHI), used to harvest energy from structural vibrations, was presented in [79]. This technique has been shown to increase the energy harvested by a factor of up to 10. In this article, SSHI was combined with

2.4. OTHER ASPECTS

27

commercially available low-energy devices to develop an autonomous wireless sensor. Among the conclusions, it was stated that many consideration had to be made regarding the specific problem at hand to achieve optimal energy savings. Thus, the design of ultra-low-energy devices was very application-specific.

Jang et al. [80] presented the results from a wireless monitoring system deployed on a large-span cable-stayed bridge. The monitoring consisted of acceleration, temperature and wind dynamic measurements. The deployed nodes were provided with energy harvesting capabilities, and acceleration was measured on the deck and cables.

Rice et al. [81] presented a node developed specifically for the wireless structural health monitoring of civil structures. The computational capabilities, power consumption and communication capabilities of the sensor were presented with the relevant information for the developed sensors to use with the nodes. The software used for data integration was also presented, and real-life deployments of the proposed method were briefly discussed.

2.4.2 Sensor Placement

Sensor placement is a crucial decision to make when designing SHM systems. The nature of the most commonly used sensors in SHM makes repositioning them very difficult and sometimes practically impossible once the monitored structure has been opened to traffic. At the same time, good measuring devices and analysis algorithms can completely fail to achieve damage identification if the sensors are uncritically placed. In many studies of new monitoring methods, sensor placement is discussed only briefly. This discussion is, in most of cases, rather superficial, but research does exist that is aimed purely at developing sensor placement strategies.

Figure 12. Outline of the complete automated wireless SHM cell-based system proposed in [78].


28

This type of research is usually confined to the aerospace industry, but some schemes are very general and could also be implemented in bridge monitoring. The need for redundancy in the SHM system has also been studied within this field, which is important because redundancy allows a SHM system to be of use even when one or more sensors fail.

For example, in [82] a scheme is developed for the optimisation of sensor placement to reduce false negatives (or false positives) within a certain confidence interval. The developed technique takes advantage of known failure rates. An example of a wave-propagation-based active sensing system illustrated the technique. The method developed was, nonetheless, general and not restrained to these types of sensing systems. The fact that it accounted for a priori knowledge about the probability of different outcomes could make it unsuitable for certain applications, but very advantageous for the many applications in which such knowledge exists.

In [83], the problem of finding the most informative sensor placement by maximising the Fisher information matrix and through energy approaches is considered. The methods were also compared for their ability to withstand noise contamination. The performed experiments were numerical.

2.4.3 Sensor Failure Detection

The serviceability life of sensors is limited, and is often some degrees of magnitude lower than that of the structure being monitored. Therefore, it is important to be able to recognise when a sensor fails so that the information coming from faulty sensors can be discarded and false damage detection alarms avoided. This problem is also a pattern recognition problem because the ways in which a faulty sensor differs from a healthy one might not be obvious. Often, the signals produced by faulty sensors are easy for an experienced operator to distinguish, but this capability should be programmed into the SHM system to keep the monitoring process as automated as possible.

In [84], for example, an algorithm to detect piezoelectric sensor failure in the form of detachment from the measured surface is developed. An experimental structural frame was built with adjustable clamps that could simulate different damaged conditions at the joint of the structural members. Sensors were adhered to the structure, and a different imperfection was introduced in half of them so that each instrumented section would have one healthy and one faulty sensor (at opposite sides). The monitoring system was then tested under different temperatures ranging from 15 to 65 °C. Information on the susceptibility of the sensor in healthy and damaged conditions was used as a base for the future identification of sensor failure. The algorithm successfully discriminated between faulty and healthy sensors, even in temperatures outside the range tested during the susceptibility study. By obviating the sensors detected as damaged and using the information provided by the healthy sensors, the developed algorithm detected damage in the form of bolt force loss simulated by a reduction in the normal force applied by the adjustable clamps. It was noted that temperature variations could produce larger changes in the susceptibility of the sensors than some forms of sensor failure. Therefore, the temperature effects should be studied for each structure-sensor system, which could be impractical in many applications.

2.5. CONCLUDING REMARKS

29

2.5 Concluding Remarks

These are some of the conclusion pertinent to recent developments in SHM and damage detection techniques and applications for bridge structures. The findings from this review are summarised below. SHM in the field of bridge engineering is still in a developing stage. New sensing technologies and methodologies for the analysis of gathered data are constantly being introduced, and most of these advances have originated in fields other than bridge engineering. Each field has its own challenges that are similar in many senses to those of bridge engineering. Nonetheless, the adaptations to real bridge structures are far from direct. Bridges are a very expensive and critical part of infrastructure, so they are built with high safety factors. As a consequence of these safety considerations, the failure of a bridge is an infrequent event that can take a long time to develop. Most SHM strategies are therefore validated only by numerical and laboratory experiments, and only a few are deployed and implemented as long-term monitoring strategies. Furthermore, even if a monitoring system is deployed, damage is not expected to appear in the short term, so the damage identification capabilities remain tested only under laboratory conditions. There is also very little documentation on the maintenance and operation costs that an in-field monitoring system will incur. The life expectancy of electronic components is nowhere near the design lives of bridges (which often surpass 100 years) and generally requires much more service than a bridge structure. There are a number of practical problems to be solved when deploying a system on a real bridge. An example of unexpected issues that the authors of this study have encountered include mice gnawing on the instrumentation cables located inside the bridge girder and sensors and amplifiers being eliminated/damaged during thunderstorms. These kind of practical problems never arise in laboratory experiments, but must be considered in a real deployment. The authors’ review of reported damage to Swedish bridges revealed that corrosion is the most common problem affecting bridges. Corrosion does not immediately bring a reduction in bending stiffness, and it only manifests as changes to the electrochemical properties of the material. Visual signs of corrosion appear only if it is superficial or very severe. A monitoring system aimed at successfully monitoring corrosion must detect it before the corrosion manifests itself as a reduction in stiffness. Sensors that are embedded within a concrete section give researchers much more freedom to choose the point at which measurements are taken, allowing direct measurement of the points of interest without the need to extrapolate data from other points. This specificity is especially important in highly localised phenomena such as corrosion. However, cast-in-place sensors are impossible to replace, service or relocate, making the monitoring system much less flexible. The amount of data recorded by a continuous monitoring system dynamically measuring a number of variables will rapidly become too large to handle. Therefore, it is preferable to immediately analyse the data locally and store only the processed results. Communication


30

with a remote computing unit is often limited, especially in more remote bridges, so data analysis must be performed locally. More research and development work will be needed before SHM can become a practical method of cutting maintenance and inspection costs in bridge structures and reduce the risk of sudden failure. SHM should, in its current state, not replace traditional inspections, but act as a complementary source of information on structural health. Intensive research is currently aimed at achieving that goal. With the new advances in sensing technologies and data analysis techniques, SHM is a promising tool that will gradually lead to reduced bridge management costs and increased safety. SHM will also definitely help us to increase our knowledge and understanding of the static and dynamic behaviour of bridge structures.

2.5. CONCLUSIONS

31

3 Bridge Weigh-In-Motion

A bridge weigh-in-motion (BWIM) system extracts information about the vehicles crossing a bridge from physical quantities measured on the bridge. The measured quantity is almost always strain because this quality is easy to measure and well-defined for a given quasi-static load. The latter fact is essential because, among all vehicle parameters, the static weight (of each axle or of the whole vehicle) is by far the most commonly sought. A BWIM consists of a bridge instrumented with sensors, a data acquisition system, an algorithm to extract the relevant vehicle parameters and a database to store the results. It is also common (especially in earlier implementations) to use axle detectors in highway implementations of BWIM to obtain vehicle speed and axle spacing. In later implementation, the axle configuration is often extracted from the strain measurements.

One of the main purposes of many WIM implementations is the enforcement of maximum load regulations, highlighting the importance of obtaining accurate static load estimations. Vehicles loaded beyond the allowable maximum have contributed to the rapid deterioration of tracks and roads. BWIM is used in several European countries as a pre-screen to select vehicles suspected of overloading that will be measured with more accurate static scales. It has been estimated that overloaded trucks have caused approximately 20% of the costs associated with pavement maintenance [85].

BWIM can also provide useful traffic characterisation. Information about the real loads affecting a given structure helps researchers understand the damage mechanisms and improve maintenance. The information obtained by a WIM system can be used, for example, to obtain the extreme cases composing the tail of the load distribution. This information is interesting in that it can help characterise the aggressiveness of the traffic. The general layout of the entire traffic distribution can also be obtained. Surveying the number of vehicles and their gross weight, speed and axle spacing is of interest to a number of disciplines.

Traffic information can help reduce costs in both assessment and design, especially for bridges. Traffic load models from bridge design codes are very conservative compared to what the actual load can be expected to be. BWIM can also be used to identify the dynamic characteristics of the traffic at a given site.

BWIM systems are preferable to pavement WIM because they register the loads for a longer time, allowing the dynamics of the signal to average out. Pavement sensors are too short and measure axle load during only a fraction of the tyre-vehicle system’s vibration period, leading to large possible deviations above or below the static value due to the vehicle’s dynamics.

Another advantage of BWIM is that it is relatively hidden compared to other weighing stations and pavement WIM. It has been demonstrated that more overloaded trucks cross roads that are not known to be instrumented [86] because BWIM systems are not readily visible to passers-by.

Chapter 3. Bridge Weigh-In-Motion

32

In general, BWIM systems are less expensive than weighing stations. Strain gauges are very inexpensive and the computational power required is well within reach of an off-the-shelf, low-priced home computer. In some BWIM applications, for example, the deployment cost can also be very low when the strain gauges perform their measurements at the soffit of the bridge.

3.1 History of BWIM

The first BWIM was developed by Moses for the US Federal Highway Administration in 1979 [87]. Moses used the fact that the strains at point A caused by a load at a certain position B will equal the product of the load magnitude and the influence line of the strains at A evaluated at point B. The influence line of a given quantity is a function that returns the value of that quantity for a given load position. By using axle detection strips, an accurate estimation of the axle position at each time can be obtained, and therefore the strain signal can be decomposed into each of the axles’ contributions. In this implementation, the strains were measured at the midspan on the soffit. The system was not originally fully automated; rather, a switch had to be manually activated when a vehicle was approaching.

This system was improved by Moses and Ghosn [88]. The enhanced system could achieve an error boundary under 6% of the gross vehicle weight, although the weight of each particular axle was estimated poorly. With this enhanced system, Moses tried to tackle the problem of multi-vehicle events.

Later, in 1984, AXWAY was developed by [89]. It calculated the gross weight of a vehicle under the assumption that the gross weight will be proportional to the area under the strain time-history. Once the gross vehicle weight had been obtained, the weight of each particular axle was determined by minimising the difference between the measured and expected signals. The error was reported to be under 3% for gross vehicle weight and 10% for axle weight in 90% of the cases. The bridge used for validating the method had large dynamic effects, mainly due to the bridge’s first eigenmode. Integrating the strain signal for a full number of periods of the bridge’s first mode of vibration was proposed as a solution, but this technique introduced inaccuracies in the calculations for closely spaced axes or very high speeds. The system required an operator because real time analysis of the data was beyond the capability of the computers available at that time. Two years later, the CULWAY system was developed [90] using culvert bridges. The dynamics of culvert bridges are much lower than those of other types of bridges, due mainly to the high damping introduced by the surrounding soil. The system included two axle detectors, one in the middle of the span and one before the bridge. The system first measured the strain at the centre of the culvert when the vehicle was still approaching the bridge, then read the strains as each axle crossed the midspan section. The axle weight was then calculated from the differences between the loaded readings and the initial unloaded reading.

Since these original algorithms, the number of new implementations has grown immensely, improving the accuracy of the methods and tackling new challenges such as including an extension to 2D [91], the load estimation from reaction force measurements [92] and the identification of dynamic loads. Interest in obtaining the actual dynamic contact force (as opposed to static weight) appeared early in the history of BWIM [93]. This problem is referred to as moving force identification (MFI). The dynamic contact force is important because it is the actual load affecting the bridge and can be used to estimate dynamic

3.2. RECENT ALGORITHMS AND APPLICATIONS

33

amplification factors. MFI is a badly-posed problem that presents serious difficulties, but major advances have been made, especially since the introduction of regularisation schemes.

The theoretical advances in WIM, and especially in moving force identification, apply in principle to any structure. In practice, however, until relatively recently, WIM algorithms had been applied only to highway bridges. This development is somewhat surprising because the maintenance costs of railway bridges can also be expected to depend on the load history, and statistics on railway traffic can also provide information useful for planning.

The first implementation of a BWIM for railway traffic was found in Liljencrantz et al. [94], who used strain measurements in an integral bridge to calculate bogie weights, bogie distances and the speed and acceleration of trains. The more or less global influence line (IL) of this particular bridge did not allow for axle detection and was instead limited to complete bogies. The system had an auto-calibration procedure that recognised known locomotives crossing the bridge from the time signal, and used the known weight and axle spacing of these locomotives to further improve the IL, reducing the estimation errors.

To the author’s best knowledge, the only other implementation of BWIM for railway bridges since Liljencranz has been Pimentel et al. [95]. This study instrumented a short-span concrete bridge carrying high-speed railway traffic. The bridge consisted of 6 simply supported spans and the instrumentation consisted of Bragg fibre optic sensors that measured strain at 2 consecutive midspans. The train speed and axle spacing were obtained from measurements performed directly on the rails because the IL of the bridge proved too long to allow for axle detection. The strain signal was filtered at 6 Hz to eliminate the contribution of the bridge eigenmodes, and the influence line was obtained using the matrix method with a train of known configuration at reduced speed.

3.2 Recent Algorithms and Applications

Kim et al. [96] introduced a new algorithm for extracting the influence line and the load position and magnitude from the time signals of strains measured in the deck plate of a bridge. The algorithm is based on the theory of optimal linear filters. The overall layout of the method proposed is very similar to Moses’ original algorithm, but is a clear improvement thereon. The influence line must be known before vehicle loads can be estimated. The influence line can be calculated as an optimal linear filter that, applied to the autocorrelation of the wheel load signal, gives the cross-correlation between the measured time signal and the wheel load signal. With the influence line obtained, the axle load and positions can be obtained using a similar procedure. The same objective equation used to obtain the influence line can be used (minimising with respect to the wheel sequence this time) to obtain the axle position and loading. If the number of data points is equal to the number of loads, this method yields Moses’ algorithm. The method depicted was used in a real bridge and its sensitivity to a number of parameters was studied. It was found to give very good results resulting from the highly local deck plate response. This influence line-finding algorithm proved robust and the load estimation accurate (within 10% of the real value).

Wang and Qu [97] introduced the concept of dynamic influence lines to calculate axle load even when the effects of bridge dynamics were relatively large. The dynamic influence line corresponds to the measured response when a unit load crosses the bridge at a certain speed. The dynamics of the bridge are included in the method, so the influence line will have a free-


34

vibration component after the train has left the bridge. Although not mentioned in the article, the dynamic influence line is speed-dependent, so it needs to be calculated for each speed. Once calculated, the total response can be obtained by a weighted combination of time-shifted dynamic influence lines in accordance with the superposition principle. The axle loads and spacing are obtained by minimising the difference between the measured and expected signal via an annealing genetic algorithm. The entire set-up was validated via numerical experiments and a number of parameters were studied to observe the method’s sensibility. It was noted that the dynamic influence line could not be obtained from measurements, and therefore a sufficiently accurate model of the structure’s modes would be required to obtain the influence line.

Deng and Cai [98] introduced a new method that separates the total response of the bridge into an inertial component, due only to the momentum of the bridge, and an interaction component, due to the vehicles moving on the bridge (damping forces are ignored). The vehicles are assumed to be single-DOF mass-spring-dashpot systems. The inertial component is calculated from the acceleration of the bridge deck and subtracted from the total response to obtain the force-induced response. The interaction forces are then calculated from the force-induced response of the bridge and the influence line (or surface). The algorithm was tested in a numerical model considering surface roughness and signal noise. The study concluded that not considering the inertial effects (usually disregarded in other MFI application) introduced unacceptable errors to the load estimation, while the proposed algorithm remained accurate and independent of vehicle speed, road surface condition and moderate levels of noise. In a follow-up paper [99], this method was tested in a real simply supported concrete bridge. The bridge was modelled numerically to obtain the influence surface and the model was updated to better fit the measurements taken for the structure. A test truck was driven on the bridge at different speeds, with and without artificial bumps in the road, to study the effect of these variables on the algorithm’s accuracy. To obtain the force-induced response, the acceleration time histories of the numerical model were used instead because it was not possible to measure the real acceleration at all points in the structure. The static value of the axle force was found to be very accurate when the load was far from the ends of the bridge, where the influence surface was close to zero. This was the case regardless of vehicle speed or surface roughness.

Lechner et al. [100] used a method originally introduced by [101] to make an accurate extraction of axle timing using a wavelet decomposition of the time signal of crack displacement. Proper estimation of the axle timing is paramount to obtaining the vehicle speed, and thus the axle spacing, without the aid of axle detectors placed on the surface or other speed-measuring instruments. Good axle timing can be hard to achieve from the time domain in bridges with broad influence lines, especially when loaded with fast-moving vehicles with short axle distances. Inaccurate estimations of the axle spacing introduce unacceptable errors in almost any BWIM. With accurate axle spacing, the axle load is then calculated via an optimisation procedure. This set-up was tested in two structures and found to give accurate gross weight estimations, but the individual axle estimations were far less accurate. In general, it was found that the bridge’s dynamic behaviour made accurate axle estimation difficult to achieve. When the dynamics were low, the results were satisfactory.

Yamaguchi et al. [102] described the implementation of BWIM in a curved bridge with skew. This BWIM was installed mainly to obtain information on the area’s heavy truck traffic. The bridge geometry made it far from ideal for BWIM, but it was the only structure available in the area. The axle loads were obtained from a modification of Moses’ algorithm using local strain influence lines. These influence lines were obtained by running a truck of known

3.2. RECENT ALGORITHMS AND APPLICATIONS

35

configuration and minimising the difference between the expected and measured results. Measurements were made in all 5 longitudinal beams at different cross sections, allowing for the extraction of vehicle speed, lane and axle configuration. The algorithm was capable of calculating the gross weight within 10%, even when multiple vehicles were present on the bridge.

Wall et al. [103] proposed a Bridge Weigh-In-Motion algorithm based on the derivatives of the strain time signal. For a simply supported beam, the second derivative of the strain measured at the central cross section will have a negative peak when the load crosses that cross section and positive peaks when the load enters and leaves the bridge, allowing for a good estimation of the axle timing. In this way, the vehicle speed and configuration can be determined. The gross vehicle weight is calculated from the area under the strain signal, which is in principle proportional to the gross vehicle weight up to a factor depending linearly on the vehicle speed. The gross vehicle weight is then distributed among the detected axles based on the fact that the peak registered in the second derivative of the strain is proportional to the magnitude of the axle load crossing that section. This method was validated on a real structure using a test load of known magnitude and configuration. The gross vehicle weight measured by this system was reported to be accurate within 7%, while the individual axle loads were inaccurately detected.

Obrien et al. [104] proposed a regularised solution to the BWIM equations that follow from Moses’ algorithm. The inverse problem that arises from Moses’ algorithm is generally ill-defined and sensitive to noise. Therefore, a regularisation method was introduced to control the conditioning of the problem. A method of obtaining the optimal regularisation parameter was also proposed and tested in a numerical experiment. It was found that the regularised solution reduced the error dramatically in the presence of noise and halved the errors compared to the non-regularised solution when sinusoidal loads were considered. Lastly, a spring-mass model was run on the bridge model. The regularised solution’s accuracy was improved by a factor of 4 when detecting single axles on smooth surfaces and by a factor 10 on rough surfaces.

Kim et al. [105] used two artificial neural networks as a pattern recognition technique to estimate vehicle characteristics from the time signals recorded in two types of bridges, a simply supported concrete bridge and a cable-stayed bridge. The results were compared to other BWIM techniques. A number of test loads were run over the bridge at different speeds repeatedly. The ANN was trained using the data acquired by the other two systems while the bridge was open to traffic. The first neural network calculates the gross vehicle weight, while the second assigns an axle weight distribution factor to each axle to obtain axle loads. The gross weight is calculated from the peak values of the strains at the main girders and/or cross beams plus the axle base and the vehicle speed, utilising 2 hidden layers. The individual axle loads are calculated from the axle distances, the peak strain values of the deck and the gross vehicle weight using 2 hidden layers. The trained neural network showed a discrepancy of less than 20% when compared to the other systems. When compared with the test load, the gross vehicle weight results were accurate within 10%, while the individual axle weights were accurate within 15%.

Dowling et al. [106] presented a method to update a finite element model for use in a moving force identification problem. The updating was performed using strain data gathered during the passage of a test vehicle of known characteristics. The proposed method was tested in numerical experiments and showed good computational efficiency, even for models with a large number of updated parameters. A moving force algorithm that required the mass and


36

stiffness matrices of the structure as input was used on numerically generated data both with the exact matrices (used to generate the data) and with the matrices obtained by the proposed updating algorithm. The results showed very good agreement.

Vala et al. [107] proposed a genetic algorithm to solve the inverse problem inherent to BWIM. They use numerically simulated data from a simply supported Timoshenko beam crossed by a two-axle vehicle modelled as moving forces with a static component and added sinusoidal components. The proposed genetic algorithm considers several species and migration to further enhance its robustness, and was shown to perform well at identifying both the magnitude and spacing of the load (3% error with noise free data and 5% error for high levels of noise, if several instrumented cross-sections were used).

Cheng and Yang [108] presented the installation, calibration and results of a BWIM system deployed in a steel truss bridge. In this deployment, the truss suspenders, transversal beams and longitudinal ribs were instrumented. The system was able to estimate gross vehicle weight, vehicle speed and lateral vehicle position as they crossed the bridge. Axle loads were also estimated with the aid of a camera to precisely time the passage of each axle. No verification of these results was offered.

Deng and Cai [109] implemented a BWIM in a numerical simply supported slab bridge and were able to perform MFI based on strain and deflection measurements. The numerical simulation was performed considering the bridge-vehicle interaction using a sprung mass system with road roughness. They studied the effect of sensor placement, lateral location of the loads, vehicle speed, number of vehicles and noise levels. In a companion paper, they applied the proposed algorithm to a real structure, studying the capabilities of the BWIM system using strain and deflection measurements and investigating the dynamic amplification factor with and without a 25-mm road bump used to increase the roughness of the road.

Edalatmanesh and Newhook [110] proposed an optimisation-based algorithm to solve the badly-posed inverse problem associated with BWIM and load identification. The suggested approach was tested in numerical and laboratory experiments. The laboratory experiment consisted of a 1:3 scale model of a steel-concrete composite bridge. A statistical study of the identification error was also presented, and it was shown that by using multiple estimations of axle loads, the error in the statistical description of the loads crossing the bridge could be reduced. The method was based on the influence matrix (a collection of influence lines for different points) and the effects of imperfect influence matrices on load identification were also studied and presented.

4.1. ARTIFICIAL NEURAL NETWORKS

37

4 Advanced Techniques for System Identification and Damage Detection

This chapter presents some of the mathematical tools used throughout this work. The presentation aims at an intuitive and non-rigorous understanding of the principles behind these tools. For more rigorous descriptions, the reader is directed to the articles attached in the appendices and the references therein.

4.1 Artificial Neural Networks

Artificial Neural Networks (ANNs) are a class of functions inspired by the way animal nervous systems work [111]. An ANN consists of a number of interconnected nodes, also known as neurons (in an analogy to the neurons in nervous systems). An ANN is defined by the weighing factors of each neuron-to-neuron connection, plus a number for each neuron called the “bias” and a real valued function associated with each neuron called an “activation function” (normally a sigmoid). Each neuron has very limited computational capabilities, taking as inputs the outputs of neurons that are “upstream” and connected to it, multiplied by the weighing factor of that specific connection. The weighted inputs and the neuron’s bias are summed together and the activation function is evaluated for this sum. The value thus obtained is the output of that neuron. This unique output is then taken as input for all of the neurons “downstream” connected to the evaluated neuron.

In the simplest case, ANNs are connected in ordered layers, with each layer taking input from the previous layer and its output being used as input for the next layer. There are no connections between neurons in the same layer and no connections flowing “backwards” in the layers. In this type of configuration, the first layer, which does not take input from any other layer, is the input layer and the last layer is considered the output layer. All of the intermediate layers, if present, are called the hidden layers, as seen in figure 13.

ANNs are used as black-box models for function approximations, clustering, pattern recognition and other proposes. They are a very popular type of machine learning algorithm. Their popularity comes from the fact that they are conceptually easy to grasp and because there are computationally efficient methods with which to train ANNs (at least those with only few layers) due to the invention of back-propagation [112].

To use an ANN as a function approximator, it must be trained first, which requires a “training set”. This training set consists of a set of input-output pairs. The weights and bias that conform to the ANN are then adjusted by a suitable algorithm so that the output of the ANN matches as closely as possible to that of the underlying function for a given input. It is important that the training data span the entire domain of possible input values.

Chapter 4. Advanced Techniques for System Identification and Damage Detection

38

Figure 13: An example of a simple ANN with a single hidden layer.

Normally, the “network architecture” (i.e., the size and arrangement of the layers and the chosen activation functions) is not changed during training, and it is up to the user to make suitable choices. The success or failure of an ANN as a function approximator depends critically on the architecture chosen [113].

Figure 13 shows a typical ANN in a layered configuration, with 2 input neurons, 3 hidden neurons and 1 output neuron. The weighting factors (w) of the connections and the biases (b) of each neuron are explicitly shown with the activation function for each neuron (sigmoid for the hidden layer and linear for the output layer). It has been shown that, given enough neurons in the hidden layer, this 3-layer configuration can approximate any function to an arbitrary degree of accuracy [114].

4.2 Gaussian Processes

Gaussian processes are a type of stochastic process, and like ANNs, they can be used in regression and classification problems. A stochastic process can be thought of as a collection of probability distributions that varies with some independent variable (which could be time, space or some other variable). A simple example of a stochastic process is a normal 6-faced die equipped with a mechanism that moves its centre of mass around continuously in a pre-set

4.2. GAUSSIAN PROCESSES

39

pattern. At any point in time, there is a well-defined probability of getting a certain number if the die is rolled. However, this probability depends on the location of the centre of mass, which changes with time (the independent variable). Therefore, the probability of obtaining a given number is a well-defined, but time-dependent, quantity.

A Gaussian process is a particular case of a stochastic process in which the probabilities are Gaussian at every point in time (or other independent variable), even if the value of the mean and standard deviation vary continuously with time.

A stochastic process can be useful to fit (regression) a function that is inherently noisy. The stochastic process will then give an estimation of how the mean and variance of the underlying function vary with time (or other independent variable) [115].

Returning to the example of a die with a moving centre of mass, imagine that the centre of mass is moved slowly enough by the mechanism to allow the die to be rolled several times while the centre of mass (and thus the probabilities involved) remain essentially unchanged. It would then be possible to estimate the expected value and variance at each point in time by rolling the die repeatedly and obtaining estimates of these variables every few rolls. Of course, the actual Gaussian process fitting algorithm does not use this naïve approach, but the example gives a certain idea of the process’s nature and what it can be used for.

Fitting a Gaussian process to certain data requires some a priori assumptions on the nature of the data. Consider the data in figure 14, which can be fitted to a rapidly varying mean with very little variance (an example of overfitting), shown in red, or to a Gaussian distribution that has a constant independent variable but a very large variance, shown in blue. Both are obviously wrong and fail to characterise the true nature of the data (the coloured area represents the area of 2 standard deviations).

The process shown in figure 15 is a much more reasonable explanation of the function underlying the data. To create a more mathematically rigorous definition of what is “reasonable”, one specifies a “characteristic length” at which the mean and standard deviation may vary significantly. This knowledge must be provided prior to the Gaussian process fitting algorithm and is part of the engineers’ judgment [116].

Mean and standard deviation functions that vary too rapidly (in comparison to the defined characteristic length) are thus assigned a low likelihood. Meanwhile, mean and standard deviation functions that vary little will need very large standard deviation to encompass all of the data points. These large deviations make the slow-varying process (like the one shown in blue in figure 14) unlikely, given the data. By finding the mean and standard deviation functions with the largest likelihood, a mathematically well-defined “optimal” process can be reached.

It is common in novelty detection algorithms to identify the probability distribution of certain quantities and use future measurements of these quantities that have low likelihoods to alert the user that something unusual is happening (normally damage). A Gaussian process can be more accurate when the probabilities involved depend significantly on other measurable quantities, particularly if this dependence is large enough that just taking the average behaviour would lead to considerable noise in the data.

In paper IV, the prediction error of the proposed algorithm was found to show this type of dependence on the vehicle speed.


40

Figure 14: Two examples of improper fitting. The red process is clearly overfitted, while the blue process is clearly “underfitted”.

Figure 15: The same data as in figure 14 is now fitted with the process with the highest likelihood.

4.3. HILBERT TRANSFORM

41

In this way, a Gaussian process approach was implemented using the vehicle speed as an independent variable. Both the input (vehicle speed) and output (prediction error) variables are one-dimensional, making the approach easy to visualise.

The Gaussian process was also used in paper III, but this time as a black box model. The method proposed in this study required many thousands of serial calls to a certain function, which was very computationally expensive. To make the algorithm faster, this function was evaluated thousands of times, but in parallel. The input-and-output pairs were used to fit a Gaussian process that was employed to interpolate the function evaluation with a speed several orders of magnitude faster, but with little accuracy lost.

4.3 Hilbert Transform

This section is a brief presentation of the Hilbert transform, its advantages and disadvantages when compared to other signal analysis tools and the uses it can have in SHM applications. For a more mathematical description of the Hilbert transform, the reader is referred to paper IV and the references therein.

Data analysis is fundamental to any engineering application. Mathematical models, which are necessarily simplifications, are used to understand the complex systems that constitute reality. Data acquired in the real structure are then used to confirm the assumptions underlying these models. Paradoxically, the assumptions implicit in the data analysis method used often impose further constraints on the model. The Fourier transform, for example, because of how it is theoretically developed, is applicable only to stationary, linear signals. It remains, however, one of the most commonly used tools in the analysis of signals from bridge monitoring, despite the fact that these signals are never stationary and seldom perfectly linear. These analyses need not be wrong, but care must be taken not to misinterpret the results they provide.

More modern techniques, such as the short-time Fourier transform and wavelet analysis, have permeated SHM from other fields. These techniques are currently fairly commonplace in health monitoring and have several advantages over the standard Fourier analysis. They still have important shortcomings, however, because both techniques must generally trade resolution in the time domain for resolution in the frequency domain. These techniques also suffer from leaking effects due to the finite time duration of the windows, in short Fourier transform, and that of the basic wavelet, in the case of wavelet analysis.

Frequency, in the Fourier sense, is well-defined only for stationary signals, and is also a property of the entire signal taken over its infinite duration. Consider, for example, a signal composed of one single sinusoidal cycle with a 1-Hz frequency (starting at zero) followed by a single sinusoidal cycle, of the same amplitude and also starting at zero, but with a frequency of 2 Hz. The total duration of this signal will be 1.5 seconds, and it starts and ends at zero with no discontinuities. Now consider a signal that consists of an infinite repetition of this 1.5-s signal. This signal is obviously periodic, with a period of 1.5 s (see figure 16a). A Fourier analysis will therefore display an energy content at 0.67 Hz and its harmonic multiples. This result is not wrong, but it somehow contradicts the more intuitive idea of frequency and frequency-modulated signals. In some applications, it would be useful to have a tool that gives us some kind of instantaneous frequency. In an intuitive but clear sense, the


42

frequency content of the signal changes in time (from 1 to 2 Hz), but that is beyond what a Fourier transform can show us.

A wavelet analysis shows energy content at the right scales, and thus at the right frequencies (approximated using the wavelet central frequency). However, the spill effect will be so large that the energy content of the wavelet analysis would be virtually constant over time (see figure 16b). The visualisation it shows corresponds more to our intuition of a frequency-modulated signal (with content in the 1-Hz and 2-Hz regions), but in some sense it still fails to represent the true nature of the signal in a way that is “instinctive”. There are no clear moments when the energy is localised at only 1 Hz or at only 2 Hz.

The Hilbert transform [117], on the other hand, gives us just the intuitive picture of a frequency-modulated signal we need. Using the Hilbert transform, one can obtain a mathematically well-defined “instantaneous frequency” that, unlike frequency in the Fourier sense and scale for the wavelet analysis, does not suffer from leakage. If applied to our example signal from before, the instantaneous frequency “jumps” from 1 to 2 Hz at the instant one would expect it to (see figure 16d). This accuracy comes at a price, of course. One reason why wavelet analysis is so popular, among others, is due to its “robustness”. Almost any signal can be analysed with wavelets, while the Hilbert transform is applicable only to a certain type of signal. It is sufficient that the signal be narrow-banded for a signal to be successfully analysed with the Hilbert transform [118]. This is not a necessary condition (the signal in the example above is not narrow-banded), but narrow-banded signals are a feature of many engineering applications, so it is a useful categorisation.

In a similar fashion, the Hilbert transform also provides the instantaneous amplitude, a quantity that in not well-defined for Fourier or wavelet analysis. To achieve well-defined instantaneous frequency and amplitude, the Hilbert transform (loosely speaking) turns an oscillating, “flat”, real-valued signal into a spiralling, “3D”, complex-valued signal. Instead of oscillating around zero and crossing it multiple times, the transformed signal coils in the complex plane around the zero axis, without ever actually equalling zero. In this way, the instantaneous frequency can be defined as the angular speed at which the signal coils around the zero axis, and its instantaneous amplitude becomes simply the distance to the zero axis, or the absolute value of the signal at a given point in time. To recover the original signal from the transformed signal, one needs only to take its real part.

4.3. HILBERT TRANSFORM

43

Figure 16: a) A synthetic frequency modulated signal, b) its continuous wavelet transform and c) the instantaneous amplitude and d) frequency as obtained by the Hilbert transform.

5.1. THE SÖDERSTRÖM BRIDGE

45

5 Case Studies

Case studies were carried out as part of this thesis and include the monitoring of three different bridges: the High Coast Suspension Bridge, the Söderström Railway Bridge, and the Skidträsk Bridge, and one numerical experiment. These bridges are important structures in the Swedish transport network, but have very different natures. Taken together, the main factors affecting the responses of these bridges, including the effects of traffic, environmental conditions and damage detection, were analysed.

5.1 The Söderström Bridge

The Söderström Bridge is a ballastless steel railway bridge located in the heart of Stockholm (see Figure 17). The bridge connects Stockholm’s Central and South stations and has the highest traffic intensity in Sweden. Over 500 trains cross the bridge every day. As such, it is a vital transport link in the Stockholm region and to the rest of the country. The Söderström Bridge is built as a continuous beam with 6 spans. The northernmost and southernmost are 27 meters long, while the inner spans are 33.6 meters long. The bridge’s main beams are 2.8 meters high and run along the sides of the bridge. Cross beams support the 2 tracks of the bridge, and stringer beams are located under each of the 4 rails to provide support in between the cross beams. The bridge contains numerous secondary bracing elements.

Theoretical studies [119] concluded that the fatigue life of the bridge had been exhausted, and several fatigue cracks had been discovered by 2008. Given the strategic importance of the bridge, the Swedish Transport Administration initiated a thorough monitoring project to assess the bridge’s condition.

Although the instrumentation was performed with other aims in sight, some of the sensors showed excellent WIM capabilities. The subject of BWIM for railway traffic is rather unexplored, and the central location of the Söderström Bridge added extra value to the study because the data on the traffic crossing the bridge would include important information on the region’s traffic.

The proposed algorithms were used as a part of this study to obtain statistical characteristics of the traffic on the Söderström Bridge. The results and methods of this study are presented in paper I.

Chapter 5. Case Studies

46

Figure 17: The Söderström Bridge as seen from the west.

Paper I

This study presented a technique for railway traffic monitoring using an instrumented bridge. The developed algorithm was implemented using only a few of the sensors from the original, very comprehensive bridge structural health monitoring system. The study included a new method to calculate the influence line that did not require prior knowledge of the test vehicle’s axle spacing. The proposed BWIM traffic load estimation technique showed great capability for axle load and axle spacing estimation using only 2 strain gauges and computationally simple and efficient algorithms. The axle load accuracy was calculated to 15% with 95% confidence and the bogie load accuracy to 8% with the same confidence.

Two of the more than 40 strain gauges comprising the original monitoring system were used. The 5 originally deployed accelerometers were also considered because it was judged useful to correlate different train configurations and loads with the deck accelerations they produced.

Novel algorithms were developed to calculate the influence lines of the sensors used and for the BWIM. The system was calibrated using a locomotive with known axle loads. Data were collected for over a month of monitoring, comprising over 7500 train passages. The algorithm was able to successfully identify 97% of the trains, resulting in train configuration (axle spacing), axle loads, train speed, train acceleration, deck acceleration (peak and RMS value) at 5 points on the bridge and information on the dynamic characteristics of each bogie. To the author’s best knowledge, this was the first implementation of BWIM for railway traffic capable of detecting single axle loads instead of bogie loads.

The algorithm easily detected wheel defects, such as wheel flats, by analysing the high frequency content of the signals. It was observed that the largest axle loads were usually associated with low dynamic amplification factors. The obtained histograms of the axle loads were best fitted by a three-modal Gaussian distribution, which could be attributed to the passenger wagon, locomotive and goods wagon. The right-hand tail of the distribution was believed to be accurately represented because the trains for which the load estimation algorithm failed generally had low axle loads.

5.2. THE SKIDTRÄSK BRIDGE

47

Thus, the developed algorithm provided useful information that could be used by bridge engineers (e.g., for bridge rating and remaining fatigue life estimation), by traffic engineers (e.g., for traffic simulations) and by railway managers (e.g., for train wheel defect detection).

5.2 The Skidträsk Bridge

The Skidträsk Bridge is a simply-supported, single-track ballasted railway bridge with a span of 36 meters (see figure 18). Two steel I-beams in composite action with the concrete slab carry loads to the supports. The beams have a reduced height at midspan, which gives the illusion of a slight arch solely for aesthetic purposes. The other dimensions (flange width and cross-sectional thickness) are larger at the midspan, so that the total effect is an increased moment of inertia at midspan.

The bridge is instrumented with accelerometers at its midspan and quarter points and with strain gauges in the concrete deck and steel beams. Some of the strain gauges measure in the transversal direction, providing a more local measurement suitable for BWIM.

The strain measurements obtained from the transversal gauges on the concrete slab make it possible to discern bogies (but not individual axles), allowing for the identification of the train type and speed.

The bridge is located in the 65th parallel, less than 2 degrees from the Arctic circle, so it experiences very low temperatures in winter. The temperature fluctuates over 60 °C annually (from -30 to 30 °C).

The bridge is crossed by regular passenger traffic and by the so-called Steel Arrow train, which transports iron ore and has axle loads of 25 tonnes. The studies performed on the Skidträsk Bridge are presented in papers II and III.

Paper II

This study addressed the very large yearly frequency shift found in the Skidträsk Bridge. The observed torsional frequency, for example, increased more than 35% from the warm to the cold season. This change was explained via a stochastic model updating scheme. Data corresponding to a year of monitoring were analysed for this updating procedure. From these data, two first eigenfrequencies were extracted for the free vibrations after each train passage and then used to update a FE model. The nature and magnitude of the observed frequency variation have not been described in the literature, to the authors’ best knowledge. It was concluded that the variation was due to the freezing of the ballast and surrounding soil, which increased the bridge’s stiffness and added major variability to the structure’s behaviour. During the cold months, the torsional frequency could vary by up to 10%. The changes were found to be linked to, but not directly correlated with, temperature. During the cold months, the mean stiffness of the soil was found to increase by 50%, while the mean stiffness of the ballast increased seven-fold.

It was believed that this phenomenon could affect other similarly ballasted bridges in cold climates and was therefore of importance to SHM, fatigue life assessment and model calibration.


48

Figure 18: The Skidträsk Bridge.

Paper III

During the studies that led to the creation of Paper II, it was discovered that, despite being a relatively simple structure, the Skidträsk Bridge displayed important nonlinearities. Paper III quantified these nonlinearities in the form of amplitude dependence of the eigenfrequencies and damping of the two first eigenmodes. This quantification was performed for both the warm and cold seasons, in which the bridge is known to behave very differently. The Hilbert transform was chosen to analyse the nonlinearities because it was proven to be more effective than the continuous wavelet transform and provided more insight into the nature of the nonlinearities. It was found that the eigenfrequencies and damping ratios were higher during the winter, and that their dependence on amplitude increased compared to the warm months. As in Paper II, it was found that all of the studied quantities showed much larger variability during the cold season.

5.3 The High Coast Bridge

The High Coast Suspension Bridge is an 1800-meter-long highway bridge located in the middle of Sweden (see figure 19). Its girder is a single continuous steel box from abutment to abutment without any intermediate vertical support at the pylons. The bridge is supported on sliding bearings at the abutments, and its length experiences a total winter-to-summer elongation of over one metre (i.e., the deck moves over 50 cm at each bearing). Because of this movement, large expansion joints are required at each end of the bridge. Large hydraulic dampers have been installed at the abutments to mitigate the effects of braking forces.

5.3. THE HIGH COAST BRIDGE

49

The sliding bearings of the bridge are made of Teflon. During a regular inspection, it was discovered that the Teflon layer had peeled off in large flakes and that the wear was far greater than expected. A monitoring programme was initiated to study the cause of this problem. The monitored variables were temperature, horizontal displacement at the northern abutment, bearing forces and the acceleration of the two hangers closer to the northern abutment. A camera was also installed to monitor the traffic passing over the bridge.

It was suspected that overloaded trucks causing large dynamic displacements and impact loading were the cause of the wear. The possibility of connecting high bearing loads with hanger forces (via eigenfrequencies), photos of the vehicles crossing the bridge, temperature variations and the displacements they produced were of great value to understanding the behaviour of the bridge and the causes behind the excess wear registered.

The system did not include a camera to monitor traffic during the early stage of the monitoring process. It was therefore impossible to determine if some instances of overloading were due to single trucks that did not comply with regulations or to a group of trucks crossing at the same time (the bridge has 4 lanes, two in each direction).

All of the measured variables were dynamically monitored at a frequency of 100 Hz, except for the photos that were taken in 4-second intervals. The amount of data accumulated during the year-long monitoring phase was therefore very large, and an automated system with which to efficiently handle the data had to be created from scratch and tailored to specifically meet the constraints and goals of this monitoring project.

A dynamic calibration measurement was carried out as part of the project. The calibration consisted of driving a truck with a known weight at different speeds on different lanes, in both directions, to determine how the loads from vehicles in different lanes were carried to each of the bearings.

Paper V summarises and presents the setup and results of the last phase of this monitoring campaign.

Paper V

From the case study, it was concluded that the monitoring system worked satisfactorily. The data gathering, developed monitoring and analysis algorithms, and the camera worked together seamlessly.

The high quality of the signals enabled the evaluation of the gross weight of vehicles passing over the instrumented bearings. This rather simple instrumentation has effectively converted this bridge into a scale for weighing all passing traffic, a so-called BWIM system.

During the measurement period, relatively large loads from traffic were registered on the bearings. The total loads acting on the bearings were still much smaller than the total static design load according to the Swedish bridge design code. Some of the large loads registered on the bearings were caused by single vehicle events (overloading, such as trucks with a gross weight of over 100 tonnes, has been measured). No significant changes could be observed in the bearing forces or the hanger forces due to the yearly temperature fluctuations.

In periods of rapid temperature change, the thermal expansion is incapable of overcoming the friction at the bearing. As a result, stresses build up and are released when a vehicle passes over the bridge. This issue could be a possible cause of the wear observed in the bearings.


50

Figure 19: The High Coast Bridge.

At the end of the monitoring project, the static loads acting on the bearings were not found to ever exceed the dimensioning values, but several overloaded trucks were in fact registered, some giving large dynamic impacts.

Lastly, it was concluded that this monitoring project and the results from the signal analysis had increased our understanding of the loads acting on the bearings, the traffic events, the type of traffic passing over the bridge and the dynamic displacements of the bridge girder. Unfortunately, we were unable to definitively identify the main cause the Teflon bearing layers’ severe wear. Therefore, we could not rule out the hypothesis that this wear was caused by the high loads and dynamic effects acting on the bearings in combination with the rapid longitudinal movements of the bridge girder on the bearings.

5.4 Damage detection

One of the case studies addressed damage detection and was based on a numerical experiment presented in paper IV.

Paper IV

This study presented a novel damage detection algorithm for railway bridges. The method was tested on a numerically simulated bridge, and the proposed algorithm used the monitored deck accelerations, together with BWIM data, to train an ANN. The trained network learned to predict the deck accelerations based on BWIM data and previous acceleration measurements (for a single specified train type). The predicted acceleration was then

5.4. DAMAGE DETECTION

51

compared stochastically with the simulated acceleration. Changes in the accuracy of the prediction were interpreted as changes in the structure and thus related to damage. The stochastic nature of the algorithm allowed it to achieve higher levels of accuracy with repeated measurement. The proposed method was model-free, which avoided the need to rely on complex FE models. Rather, the method required only unsupervised training, making it easy to apply to existing structures. The proposed method was shown to be successful at detecting damage in a simply supported beam, but should be tested in more complex structures to evaluate its performance.

6.1. GENERAL CONCLUSIONS

53

6 Conclusions

This chapter presents some general conclusions and suggestions for future work based on the work presented in this thesis. More detailed conclusions on each of the case studies are presented in the corresponding papers.

6.1 General Conclusions

It is a well-known fact that normal environmental conditions can cause changes in the behaviour of bridges that are far larger than what acceptable levels of damage can cause. The studies performed showed that the environmentally induced changes were even larger than previously thought for some bridges, making the study of these changes an absolute necessity if any damage detection is to be attempted. Even at a given temperature and load case, the responses of the structure could vary wildly due to certain inertial effects of the environmental conditions. That is, the behaviour of the structure was influenced by the environmental conditions not only at the current instant but also by the conditions in past hours.

It was also shown that relatively simple bridge structures present nonlinearities with measurable effects, even in the relatively narrow range of amplitudes present during the free vibrations. This result raised the question of how valid live load simulations can be, when the vibrations amplitudes involved are on the order of some m/s2, and the model used to performed this simulations are calibrated using on data from free vibrations with amplitudes orders of magnitude lower.

BWIM was found to give very good results for railway bridges (very good accuracy in vehicle and speed identification and good axle load identification), especially when compared to the use of BWIM for highway traffic. The limited configurations of vehicles, long unloaded state periods between train passages, total absence of variation in the lateral positioning of the loads and large numbers of axles in each vehicle all contributed to easing the identification of vehicle speed and acceleration, and axle configuration. Real-time, axle-by-axle information on the actual loads affecting a structure can be of enormous importance when rating a structure or considering its remaining fatigue life, to name just two examples. There are also many other application areas, such as the detection of wheel defects, besides bridge monitoring. The author was very surprised to see how little had been performed in the field of railway BWIM, which is definitely an underused tool within the field of bridge monitoring.

As the price of monitoring equipment and computational power decreases, the amount of data available will be overwhelming (in many senses, it is already overwhelming for some deployments). The damage detection algorithm suggested in this study was tested only numerically and in a simple structure. Nonetheless, it is the author’s belief that machine

Appendix 6. Conclusions

54

learning will play an ever more important role in damage detection as the most practical method of dealing with the information flow that will characterise the coming years. The proposed method was very successful and describes an avenue of research that should be tested further.

6.2 Further Research

Possibilities for SHM have been explored throughout this study. It is the author’s strong belief that SHM possesses great potential for more economical and rational infrastructure in general and for bridges in particular.

Although real practical applications in damage detection are far from being able to replace the traditional inspection-based maintenance plans, SHM has found niches in which it has been of great help. For example in understanding the mechanisms behind a specific problem the structure may possess or monitoring the evolution of already-present damage. However, SHM is not limited to damage detection and has very interesting branches that will be useful for engineers in the future.

An obvious future development of SHM is the continuous improvement of damage detection algorithms and sensor technologies. However, great effort will have to be dedicated to the issue of integration into bridge management systems for SHM to realise its full potential. SHM remains, in many cases, an academic pursuit. The large quantity of data generated can have a major positive impact on the way bridges are managed, but this impact requires a paradigm shift that can move bridge owners towards monitoring-based planning for inspection and management. It is an important development that bridge owners and researchers need to undertake together.

The ultimate argument that can convince a bridge owner to use SHM will be the economic argument. A small but active field of research is looking at SHM from a life cycle cost perspective. Examining the economic impact that SHM can have will be an important step towards implementing it as a mainstream tool in infrastructure management. Further research into the long term profitability of SHM is needed in order to make monitoring more attractive to bridge owners.

SHM is costly and usually requires days of work for equipment installation and removal, which limits its applicability in many cases. Wireless sensor networks have grown in importance and have started to mature to a point of being useful in SHM. Its advantages, as easy deployment, could make feasible monitoring projects that would be impossible with traditional wired technology. There are nonetheless, very few applications of wireless SHM, particularly in the medium to long term, so its advantages remain theoretical. More successful field deployments are needed to prove the value of this approach.

The many challenges in wireless monitoring, including efficient data transfer and power management, are and will continue to be a main research focus within SHM in the coming years. One of the interesting and very active research domains, that of compressive sensing, appears particularly promising. The amount of data transmitted can be reduced dramatically with compressive sensing, thus extending the battery life of wireless sensors (communication is the most power-hungry task in wireless monitoring). At the same time, this data reduction

6.2. FURTHER RESEARCH

55

is done with no significant increase in computational power, as with other compressing techniques, and it reduces the sensors’ power consumption by sampling more sparsely.

Today’s society is characterised by the ease of information flow. The Internet of Things includes not only people but also all kind of connected objects. Cars that automatically exchange information with neighbouring cars to maintain safe distances and inform of possible dangers ahead seem like a fairly inevitable future. When this technology permeates structural engineering from the automotive industry, every vehicle crossing will become a potential sensor, communicating data to bridge managers. The possibilities for SHM in this scenario are varied and rich.

If this kind of connectivity is realised, then the problem will be not one of data scarcity, but of overwhelming amounts of data. It is already the case for many bridge monitoring projects that they gather more data than can be reasonable analysed. The author therefore believes that automated algorithms that make use of machine learning, such as the one proposed in paper IV, to interpret data will be ever more important in the future.


57

Bibliography

[1] Thomson, Douglas J. "The economic case for service life extension of structures using structural health monitoring based on the delayed cost of borrowing." Journal of Civil Structural Health Monitoring 3.4 (2013): 335-340.

[2] Sedlacek G, Hechler O, Lösche A, Kühn B. "Guidelines for Monitoring of Steel Railway Bridges." Sustainable Bridges, Background document D5.2.S4; 2007.

[3] Feltrin G (Ed.). "Monitoring Guidelines for Railway Bridges." Sustainable Bridges, Report D5.2; 2007

[4] Wiber J. "Bridge Monitoring to Allow for Reliable Dynamic FE Modelling." Stockholm (Sweden): Department of Civil and Architectural Engineering, Kungliga Tekniska Högskolan; 2006.

[5] Posenato D. "Model-Free Data Interpretation for Continuous Monitoring of Complex Structures." Lausanne (Switzerland): School of Architecture, Civil and Environmental Engineering, École Polytechnique Fédérale de Lausanne; 2009.

[6] Gao JQ, Shi B, Zhang W, Zhu H. "Monitoring the stress of the post-tensioning cable using fiber optic distributed strain sensor." Measurement 2006; 39(5):420-28.

[7] Karoumi R, Wiberg J, Liljencrantz A. "Monitoring traffic loads and dynamic effects using an instrumented railway bridge." Engineering Structures 2005; 127: 1813-19.

[8] Sohn H, Farrar C, Hemez F, Shunk D, Stinemates D, Nadler B, Czarnecki J. "A Review of Structural Health Monitoring Literature: 1996-2001." Los Alamos National Laboratory Report, LA-13976-MS; 2003.

[9] Feltrin G, Meyer J, Boyle W, Gebremichael YM, Diaz A, Cruz P, Aho T, Kilpelä A, Lyöri V, Ryynänen T, Krüger M. "Monitoring Instrumentation and Techniques. Sustainable Bridges, Background document D5.1." 2007.

[10] Doebling S, Farrar C, Prime M, Shevitz D. "Damage identification and health monitoring of structural and mechanical systems from changes in their vibration characteristics: a literature review." Los Alamos national laboratory report 1996; LA-13070-MS.

[11] Salane HJ, Baldwin JW and Duffield RC. "Dynamic approach for monitoring bridge deterioration." Transport research record 1981; 832,21-28.

[12] Kato M and Shimada S (1986). "Vibration pf PC bridge during failure process." ASCE journal of structural engineering 1986; 112, 1692-1703.

Appendix 0. Bibliography

58

[13] Biswas M, Pandey AK and Samman MM. "Modified chain-code computer vision techniques for interrogation of vibration signatures for structural fault detection." Journal of sound of vibration 1990; 175(1), 89-104.

[14] Turner JD and Pretlove AJ. "A study of te spectrum of traffic-induced bridge vibration." Journal of sound and vibration 1988; 122, 31-42.

[15] Mazurek DF and DeWoIf JT. "Experimental Study of Bridge Monitoring Technique." ASCE Journal of Structural Engineering 1990; 116, 2532-2549.

[16] Spyrakos C, Chen HL, Stephens J and Govidaraj V. "Evaluating Structural Deterioration Using Dynamic Response Characterization." Proc. intelligent Structures, Elsevier Applied Science 1990; 137-1 54.

[17] Jain BK. "Diagnostics Through Experimental Vibration Signature Analysis of Prestressed Concrete ridges." International Symposium on Fracture in Steel and Concrete Structures 1991; Madras, India , 1123-1136.

[18] Tang JP and KM Leu. "Vibration Tests and Damage Detection of PIC Bridges." Journal of the Chinese Institute of Engineers 1991; 14, 531 -536.

[19] Law SS, P Waldron and C Taylor. "Damage Detection of a Reinforced Concrete Bridge Deck Using the Frequency Response Function." Proc. of the 10th lnternational Modal Analysis Conference 1992; 772-778.

[20] Raghavendrachar M and AE Aktan. "Flexibility by Multireference Impact Testing for Bridge Diagnostics." ASCE Journal of Structural Engineering 1992; 118, 2186-2203.

[21] Pandey AK, Biswas M and Samman MM. "Damage detection from changes in curvature mode shapes." Journal of Sound and Vibration 1991; 145(2), 321–332.

[22] Chance J, Tomlinson GR and Worden K. "A simplified approach to the numerical and experimental modelling of the dynamics of a cracked beam." Proc. of the 12th International Modal Analysis Conference 1994; 778–785, 1994.

[23] Aktan AE, Lee KL, Chuntavan C and Aksel T. "Modal Testing for Structural Identification and Condition Assessment of Constructed Facilities." Proc. of 12th International Modal Analysis Conference 1994; 462–468.

[24] Omenzetter P, Brownjohn JM. "Application of time series analysis for bridge monitoring." Smart materials and structures 2006; 15, 129-138.

[25] Frangopol D, Strauss A, Kim S. "Bridge Reliability assessment based on monitoring." Journal of Bridge Engineering 2008; 13(3), 258-270.

[26] Jiang R, Járegui DV, White KR. "Close-range photogrammetry applications in bridge measurement: Literature review." Measurement 2008; 41: 823-34.

[27] Huang PS, Jin F, Chiang FP. "Quantitative evaluation of corrosion by a digital fringe projection technique." Optics and Laser in Engineering 1999; 31(5): 371-380.


59

[28] Bernstone C, Heyden A. "Image analysis for monitoring of crack growth in hydropower concrete structures." Measurement 2009; 42: 878-893.

[29] Lee JJ, Shinozuka M. "A vision-based system for remote sensing of bridge displacement." NDT&E International 2006; 39: 425-431.

[30] Fu GK, Moosa AG. "Structural damage diagnosis using high resolution images." Structural Safety 2003; 23(4): 281-295.

[31] Tung PC, Hwang YR, Wu MC. "The development of a Mobile manipulator Imaging system for bridge crack inspection." Automation in Construction 2002; 11(6): 717-729.

[32] Li HN, Li DS, Song GB. "Recent applications of fiber optic sensors to Health monitoring in civil engineering." Engineering Structures 2004; 26(11): 1647-1657.

[33] Basheer M, Srinivasan S, Sun T, Grattan K, Grattan S, Taylor S. "Novel fibre optic sensors for monitoring Physical and chemical changes in concrete structures." First Middle East Conference on Smart Monitoring, Assessment and Rehabilitation of Civil Structures, Dubai, UAE, 8–10 February 2011.

[34] Bernini R, Minardo A, Zeni L. "Accurate high-resolution fiber-optic distributed strain measurement for structural health monitoring." Sensors and Actuators A: Physical 2006; 134(2): 389-395.

[35] Wu Z, Xu B, Hayashi K, Machida A. "Distributed optic fiber sensing for a full-scale PC girder strengthened with prestressed PBO sheets." Engineering Structures 2006; 28(7): 1049-1059.

[36] Majumder M, Kumar T, Kumar A, Dasgupta K, Bhattacharya DK. "Fiber Bragg gratings in structural health monitoring – Present status and applications." Sensors and Actuators 2008; 147(1): 150-164.

[37] Muralidharan S, Hyun T, Hyo J, Cheol Y, GooH, Kyeong D. "A promising potencial embeddable sensor for corrosion monitoring application in concrete structures." Measurement 2007; 40(6): 600-606.

[38] Yoo JH, Park ZT, Kima JG, Chung L. "Development of a galvanic sensor system for detecting the corrosion damage of the steel embedded in concretes structures." Cement and Concrete Research 2003; 33: 2057-2062.

[39] Sathiyanarayanan S, Natarajan P, Saravanan K, Srinivasan S, Venkatachari G. "Corrosion monitoring of steel in concrete by galvonostatic pulse technique." Cement & Concrete Composites 2006; 28(7): 630-637.

[40] Marsh PS, Frangopol DM. "Reinforced concrete bridge deck reliability model incorporating temporal and spatial variations of probabilistic corrosion rate sensor data." Reliability Engineering and System Safety 2008; 93(3): 394-409.

[41] Castellini P, Martarelli M, Tomasini EP. "Laser Doppler Vibrometry: Development of advanced solutions answering to technology’s needs." Mechanical Systems and Signal Processing 2006; 20(6): 1265-1285.


60

[42] Doebling, Scott W., et al. "Damage identification and health monitoring of structural and mechanical systems from changes in their vibration characteristics: a literature review." No. LA--13070-MS. Los Alamos National Lab., NM (United States), 1996.

[43] Carden, E. Peter, and Paul Fanning. "Vibration based condition monitoring: a review." Structural Health Monitoring 3.4 (2004): 355-377.

[44] Li, Hui, et al. "Fractal Dimension-Based Damage Detection Method for Beams with a Uniform Cross-Section." Computer-Aided Civil and Infrastructure Engineering 26.3 (2011): 190-206.

[45] Carey, Ciaran H., Eugene J. O'Brien, and Jennifer Keenahan. "Investigating the Use of Moving Force Identification Theory in Bridge Damage Detection." Key Engineering Materials 569 (2013): 215-222.

[46] Liu, T. Y., et al. "Identification and monitoring of bridge health from ambient vibration data." Journal of Vibration and Control 17.4 (2011): 589-603.

[47] Kim, Jeong-Tae, et al. "Hybrid health monitoring of prestressed concrete girder bridges by sequential vibration-impedance approaches." Engineering Structures 32.1 (2010): 115-128.

[48] Rodrigues, Carlos, et al. "Development of a long-term monitoring system based on FBG sensors applied to concrete bridges." Engineering Structures 32.8 (2010): 1993-2002.

[49] Holford KM. "Acoustic Emission in Structural Health Monitoring." Key Engineering Materials 2009; 413-414: 15-28.

[50] Lura P, Couch J, Jensen OM, Weiss J. "Early-age acoustic emission measurement in hydrating cement paste: Evidence for cavitation during solidification due to self-desiccation." Cement and Concrete Research 2009; 39(10): 861-867

[51] Carpinteri A, Lacidogna G, Pugno N. "Structural damage diagnosis and life-time assessment by acoustic emission monitoring." Engineering Fracture Mechanics 2007; 74(1-2): 273-289

[52] Hensman J, Mills R, Pierce SG, Worden K, Eaton M. "Locating acoustic emission sources in complex structures using Gaussian processes." Mechanical Systems and Signal Processing 2010; 24(1): 211-223.

[53] McLaskey GC, Glaser SD, Grosse CU. "Beamforming array techniques for acoustic emission monitoring of large concrete structures." Journal of Sound and Vibration 2009; 329: 2384-2394, doi: 10.1016/j.jsv.2008.08.037.

[54] Roberts TM, Talebzadeh M. "Acoustic emission monitoring of fatigue crack propagation." Journal of Constructional Steel Research 2003; 59(6): 695-712.

[55] Nair A and Cai CS. "Acoustic emission monitoring of bridges: Review and case studies." Engineering structures 32.6 (2010): 1704-1714.


61

[56] Shih HW, Thambiratnam DP, Chan THT. "Vibration based structural damage detection in flexural members using multi-criteria approach." Journal of Sound of Vibration 2009; 323(3-5): 645-661.

[57] Reynders E, DeRoeck G. "A local flexibility method for vibration-based damage localization and quantification." Journal of Sound and Vibration 2010; 329(12): 2367-2383, doi: 10.1016/j,jsv.2009.04.026.

[58] Kim JT, Park JH, Lee BJ. "Vibration-based damage detection monitoring in model plate-girder bridges under uncertain temperature conditions." Engineering Structures 2007; 29(7): 1345-1365.

[59] Fang SE, Perera R. "Power mode shapes for early damage detection in linear structures." Journal of Sound and Vibration 2009; 324(1-2): 40-56.

[60] Magalhães F, Cunha A, and Caetano E. "Vibration based structural health monitoring of an arch bridge: from automated OMA to damage detection." Mechanical Systems and Signal Processing 28 (2012): 212-228.

[61] Radzienski M, Krawczuk M, and Palacz M. "Improvement of damage detection methods based on experimental modal parameters." Mechanical Systems and Signal Processing 25.6 (2011): 2169-2190.

[62] Gul M, Catbas FN. "Statistical pattern recognition for Structural Health Monitoring using time series modeling: Theory and experimental verification." Mechanical Systems and Signal Processing 2009; 23(7): 2192-2204.

[63] Park HW, Sohn H. "Parameter estimation of the generalized extreme value distribution for structural health monitoring." Probabilistic Engineering Mechanics 2006; 21(4): 366-376.

[64] Gul M, and Necati CF. "Structural health monitoring and damage assessment using a novel time series analysis methodology with sensor clustering." Journal of Sound and vibration 330.6 (2011): 1196-1210.

[65] Koh BH, Dyke SJ. "Structural health monitoring for flexible bridge structures using correlation and sensitivity of modal data." Computers & Structures 2007; 85(3-4): 117-130.

[66] Jafarkhani R, and Sami FM. "Finite element model updating using evolutionary strategy for damage detection." Computer-Aided Civil and Infrastructure Engineering 26.3 (2011): 207-224.

[67] Nobahari M and Seyedpoor SM. "Structural damage detection using an efficient correlation-based index and a modified genetic algorithm." Mathematical and Computer modelling 53.9 (2011): 1798-1809.

[68] Meruane V and Heylen W. "A hybrid real genetic algorithm to detect structural damage using modal properties." Mechanical Systems and Signal Processing 25.5 (2011): 1559-1573.


62

[69] Lam HF, Ng CT. "The selection of patter features for structural damage detection using an extended Bayesian ANN algorithm." Engineering Structures 2008; 30(10): 2762-2770.

[70] Park JH, Kim JT, Hong DS, Ho DD, Yi JH. "Sequential damage detection approaches for beams using time-modal features and artificial neural networks." Journal of Sound and Vibration 2009; 323(1-2): 451-474.

[71] Reda Taha MM, Lucero J. "Damage identification for structural health monitoring using fuzzy pattern recognition." Engineering Structures 2005; 27(12): 1774-1783.

[72] Ko, J. M., et al. "Investigation concerning structural health monitoring of an instrumented cable-stayed bridge." Structures & Infrastructure Engineering 5.6 (2009): 497-513.

[73] Jiang SF, Zhang CM, and Zhang S. "Two-stage structural damage detection using fuzzy neural networks and data fusion techniques." Expert systems with applications 38.1 (2011): 511-519.

[74] Nichols JM, Seaver M, Trickey ST. "A method for detecting damage-induced nonlinearities in structures using information theory." Journal of Sound and Vibration 2006; 297(1-2): 1-16.

[75] Overbey LA, Todd MD. "Dynamic system change detection using a modification of the transfer entropy." Journal of Sound and Vibration 2009; 322(1-2): 438-453.

[76] Fan W, and Qiao P. "Vibration-based damage identification methods: a review and comparative study." Structural Health Monitoring 10.1 (2011): 83-111.

[77] Feltrin G, Saukh O, Bischoff R, Meyer J, Motavalli M. "Structural monitoring with wireless sensor network: Experiences from field deployments." First Middle East Conference on Smart Monitoring, Assessment and Rehabilitation of Civil Structures, Dubai, UAE, 8–10 February 2011.

[78] Mascarenas DDL, Flynn EB, Todd MD, Overly TG, Farinholt KM, Park G, Farrar CR. "Development of capacitance-based and impedance-based wireless sensors and sensor nodes for structural health monitoring applications." Journal of Sound and Vibration 2010; 329(12): 2410-2420

[79] Lallart M, Guyomar D, Jayet Y, Petit L, Lefeuvre E, Monnier T, Guy P, Richard C. "Synchronized switch harvesting applied to self-powered smart systems: Piezoactive microgenerators for autonomous wireless receivers." Sensors and Actuators A : Physical 2008; 147(1): 263-272.

[80] Jang S et al. "Structural health monitoring of a cable-stayed bridge using smart sensor technology: deployment and evaluation." Smart Structures and Systems 6.5-6 (2010): 439-459.

[81] Rice JA., et al. "Flexible smart sensor framework for autonomous structural health monitoring." Smart structures and Systems 6.5-6 (2010): 423-438.


63

[82] Flynn EB and Todd MD. "A Bayesian approach to optimal sensor placement for structural health monitoring with application to active sensing." Mechanical Systems and Signal Processing 2010; 24(4): 891-903, doi: 10.1016/j.ymssp.2009.09.003.

[83] Meo M, and Zumpano G. "On the optimal sensor placement techniques for a bridge structure." Engineering Structures 2005 27.10: 1488-1497.

[84] Grisso BL, Inman DJ. "Temperature corrected sensor diagnostics for impedance-based SHM." Journal of Sound of Vibration 2010; 329(12): 2323-2336, doi: 10.1016/j.jsv.2009.04.007.

[85] Henny RJ. "Experimental use of WIM with video for overloading." Proceedings of 2nd European conference on weigh-in-motion of road vehicles 1998, Lisbon, 355-363.

[86] Nowak AS, Nassif H and DeFrain L. "Effect of truck loads on bridges." Journal of transportation engineering 1993, 116(6), 853-867.

[87] Moses F. "Weigh-in-motion system using instrumented bridges." ASCE transportation engineering journal 1979, 105(3), 233-249.

[88] Moses F and Ghosn M. "Instrumentation for weighing trucks-in-motion for highway bridge loads." Case western university 1983, final report FHWA/OH83/001

[89] Peters RJ. "A system to obtain vehicle axle weighings." Proceedings 12th Australian road research board conference 1984, Hobart, 12(2), 10-18.

[90] Peters RJ. "CULWAY an unmanned and undetectable highway speed vehicle weighing system." Proceedings 13th Australian Road Research Board Conference 1986, Adelaide, 13(6), 70-83.

[91] Quilligan M, Karoumi R, and OBrien EJ. "Development and Testing of a 2-Dimensional Multi-Vehicle Bridge-WIM Algorithm." 3rd International Conference on Weigh-in-Motion 2002, Orlando, 199-208.

[92] Ojio T, Yamada K, Wakao M and Inden T. "Bridge Weigh-in-Motion using reaction force and analysis of traffic load characteristics." Journal of structural engineering 2003, 49(2), 743-753.

[93] Chan T and O'Connor C. "The use of bridges to measure dynamic wheel loads." Proceedings 10th Australian conference on mechanics of structures and materials 1986, Adelaide, 599-604.

[94] Liljencrantz A, Karoumi R and Olofsson P. "Implementing bridge weigh-in-motion for railway traffic." Computers and Structures 2006, 85, 80-88.

[95] Pimentel RM, Barbosa B, Costa NMS, Ribeiro DRF, de Almeida Ferreira LA, Araujo FMM and Calcada, RAB. "Hybrid fiber-optic/electrical measurement system for characterization of railway traffic and its effects on a short span bridge." Sensors Journal 2008, IEEE, 8(7), 1243-1249.


64

[96] Kim BH, Park MS, Park HG, Bae IH, Kim SJ and Yeo KS. "Evaluating axle loads of moving vehicle using bridge deck plate responses." Proceedings of the IMAC-XXVIII 2010, Jacksonville, USA, 1-4 February 2010.

[97] Wang Y and Qu WL. "Moving train loads identification on a continuous steel truss girder by using dynamic displacement influence line." International journal of steel structures 2011, 11(2), 109-115.

[98] Deng L and Cai CS. "Identification of dynamic vehicular axle loads: theory and simulations." Journal of vibration and control 2010, 16(14), 2167-2194.

[99] Deng L and Cai CS. "Identification of dynamic vehicular axle loads: demonstration by a field study." Journal of vibration and control 2010, 17(2), 183-195.

[100] Lechner B, Lieschnegg M and Mariani O. "A wavelet-based bridge weigh-in-motion system." International journal on smart sensing and intelligent systems 2010, 3(4), 573-591.

[101] Chatterjee P, Obrien E, Li Y, González A. "Wavelet domain analysis for identification of vehicle axles from bridge measurements." Computers and structures 2006, 84, 1792-1801.

[102] Yamaguchi E, Kawamura SI, Matuso K, Matsuki Y and Naito Y. "Bridge-weigh-in-motion by two-span continuous bridge with skew and heavy-truck flow in Fukuoka area, Japan." Advances in structural engineering 2009, 12(1), 115-125.

[103] Wall C, Christenson R, McDonnell AM and Jamalipour A. "A non-intrusive bridge weigh-in-motion system for a single span steel girder bridge using only strain measurements." Connecticut Department of Transportation 2009, report No CT-2251-3-09-5.

[104] Obrien E, Rowley C, Gonzalez A and Green M. "A regularised solution to the bridge weigh-in-motion equations." International journal heavy vehicle systems 2009, 16(3), 310-327.

[105] Kim S, Lee J, Park MS and Jo BW. "Vehicle signal analysis using artificial neural networks for a bridge weigh-in-motion system." Sensors 2009, 9(10), 7943-7956.

[106] Dowling J, OBrien EJ and González A. "Adaptation of Cross Entropy optimisation to a dynamic Bridge WIM calibration problem." Engineering Structures 2012, 44: 13-22.

[107] Vala, G., I. Flood, and E. Obonyo. "Truck Weigh-in-Motion using Reverse Modeling and Genetic Algorithms." Computing in Civil Engineering 2011. 219-226.

[108] Chen W and Yang G. "Theory and case study of vehicle load identification based on BWIM of steel truss bridge." Stahlbau 2013 82.3: 214-222.

[109] Deng L, and Cai CS. "Identification of dynamic vehicular axle loads: Theory and simulations." Journal of Vibration and Control 2010 16.14: 2167-2194.


65

[110] Edalatmanesh R and Newhook JP. "Using search based optimization algorithms in Bridge Weigh-In-Motion systems." Bridge Structures-Assessment, Design & Construction 2010 6.3: 107-119.

[111] Braspenning PJ, Thuijsman F and Weijters A, eds. "Artificial neural networks: an introduction to ANN theory and practice." Vol. 931. Springer, 1995.

[112] Rumelhart DE, Hinton GE and Williams RJ. "Learning representations by back-propagating errors." Nature 1986 323 (6088): 533–536. doi:10.1038/323533a0.

[113] Rafiq MY, Bugmann G and Easterbrook DJ. "Neural network design for engineering applications." Computers & Structures 2001 79.17: 1541-1552.

[114] Cybenko G. "Approximation by superpositions of a sigmoidal function." Mathematics of control, signals and systems 1989 2.4: 303-314.

[115] Rasmussen CE and Williams C. "Gaussian processes for machine learning." Boston: The MIT Press; 2006. ISBN 0-262-18253-X.

[116] Rasmussen, CE. "Gaussian processes in machine learning." Advanced Lectures on Machine Learning. Springer Berlin Heidelberg, 2004. 63-71.

[117] Huang NE, et al. "The empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary time series analysis." Proceedings of the Royal Society of London. Series A: Mathematical, Physical and Engineering Sciences 1998 454.1971: 903-995.

[118] Nuttall, AH. "On the quadrature approximation to the Hilbert transform of modulated signals." Proceedings of the IEEE 1966, 54(10), 1458-1459.

[119] Leander J. "Improving a bridge fatigue life prediction by monitoring." Licentiate Thesis, KTH Royal Institute of Technology 2010, Sweden.

application of monitoring to dynamic characterization and …745269/fulltext01.pdf · in a steel...

Documents